JP2004045356A - Surface defect detection method - Google Patents

Abstract

A surface defect detection method capable of accurately detecting a defect and discriminating a type by processing an image signal obtained by imaging by a novel method.
First, a gray image is obtained by imaging the surface of an object to be inspected. Next, at least a part of the obtained image signal is subjected to a top hat transform operation using one or a plurality of structural elements having a specific shading in a grayscale morphology. This makes it possible to remove the influence of the background noise and emphasize the defect to form an image. The defect is extracted from the image obtained as a result.

Description

【００６０】
またＭｉｎｋｏｗｓｋｉ差は
【００６１】
【数２】

【００６２】
と定義される。なお、Ｍｉｎｋｏｗｓｋｉ差を定義する記号には、通常○の中に−を入れた記号が使用されるが、フォントの制約上こうした記号が使用できないので、Θで代用することにする。更にｄｉｌａｔｉｏｎ（ずらし重ね）演算はＭｉｎｋｏｗｓｋｉ和のｇ（Ｘ）を対称なｇ^ｓ（Ｘ）＝ｇ（−Ｘ）で置き換えて、
【００６３】
【数３】

【００６４】
またｅｒｏｓｉｏｎ（掻き取り）演算はＭｉｎｋｏｗｓｋｉ和のｇ（Ｘ）をｇ^ｓ（Ｘ）で置き換えて、
【００６５】
【数４】

【００６６】
で定義される。関数ｆ（Ｘ）に対してｇ（Ｕ）を平行移動し、両者の垂直方向の対応点同士の和の最大値をとればｄｉｌａｔｉｏｎ演算であり、差の最小値をとればｅｒｏｓｉｏｎ演算になる。
ここでｏｐｅｎｉｎｇ演算はＭｉｎｋｏｗｓｋｉ和とｅｒｏｓｉｏｎ演算を組み合わせて
【００６７】
【数５】

【００６８】
ｃｌｏｓｉｎｇ演算はＭｉｎｋｏｗｓｋｉ差とｄｉｌａｔｉｏｎ演算を組み合わせて
【００６９】
【数６】

【００７０】
となる。
【００７１】
図１に、１次元関数ｆ（ｘ）に対して、半円状の構造関数ｇ１（ｘ）にてｏｐｅｎｉｎｇ演算を行った結果と、半円状の構造関数ｇ２（ｘ）にてｃｌｏｓｉｎｇ演算を行った結果を示す。この場合、ｏｐｅｎｉｎｇ演算は下からの平滑化処理、ｃｌｏｓｉｎｇ演算は上からの平滑処理に対応する。
【００７２】
ｏｐｅｎｉｎｇ演算による下からの平滑化処理を行った画素の値を、元画素の値からから引き算をすれば下からのトップハット変換、ｃｌｏｓｉｎｇ演算による上からの平滑処理を行った画素の値を、元画素の値からから引き算をすれば上からのトップハット変換となる。画像からの欠陥信号抽出の観点で言えば下からのトップハット変換は正極性欠陥信号の抽出、上からのトップハット変換は負極性欠陥信号の抽出を意味している。
【００７３】
上からのトップハット変換について図２で１次元信号の場合を例にとって説明する。関数の値がゆるやかに変化しているところでは、球はそれを正確になぞっていくことができるが、負のパルスがあると、その部分には球が入れないために、トレース可能な範囲からはずれることになる。よって、球の軌跡としては、（ｂ）に示すように、滑らかな曲線が得られることになる。この球の軌跡を元信号から引き算すれば鋭い負極性信号のみ抽出することができる（ｃ）。
【００７４】
図２から明らかなように、どのような信号を抽出するかは構造要素の形態によって決定される。また２次元の場合についても図３のような構造要素を用いることにより、鋭い変化を有する欠陥信号のみを抽出できることは、容易に想像できる。
【００７５】
図３においては、（ａ）に示すように、構造要素として下に凸な楕円球の形状をしたものでトップハット変換を行っている。（ｂ）が変換前の画像（グレー画像であるが、出願書類に制約のため２値化画像となっている）であり、緩やかな階調の変化の中に、鋭い階調の変化を有する２つの欠陥が存在している。実際の構造要素の大きさは、（ｂ）の下側に示される大きさである。（ｃ）がトップハット変換後のグレー画像であり、欠陥のみが抽出されていることが分かる。
【００７６】
更に真の有害欠陥の抽出をするのみならず、欠陥の種類および有害度について分類を行うとき、複数の構造要素にてトップハット変換を行うことが考えられる。以下の例では、簡単のため￣／｜＼の方向性と特定のスケールを持つ複数の構造要素を使用する場合について説明するが、これに限定されるものではなく更に複雑な形態を構造要素として採用してもかまわない。
【００７７】
このように、構造要素を用いてトップハット変換を行い、トップハット変換後の画像に基づいて欠陥の検出、程度判別、種類判別を行うのが本発明の特徴部であり、トップハット変換後の画像に基づいて欠陥の検出、程度判別、種類判別の方法は特に規定せず、従来用いられているものを適宜使用してもよいし、極端な場合には人間がトップハット変換後の画像を見てこれらのことを行ってもよい。
【００７８】
以下、欠陥の種別判定の例を説明する。最初に、欠陥が方向性を有するか、方向性のない欠陥かを判別する方法について説明する。
【００７９】
まず各構造要素にてトップハット変換後Ｓ_￣（ｘ、ｙ）、Ｓ_／（ｘ、ｙ）、Ｓ_｜（ｘ、ｙ）、Ｓ_＼（ｘ、ｙ）を得て、適当な閾値により複数の２値化を行い、各残存画素数を算出する。この例では￣／｜＼の方向性と特定のスケールを持つ４つの構造要素を使用して行い図７に示したとおり４つの画像の各残存画素数を得る。このとき複数の処理後画像Ｓ_￣（ｘ、ｙ）、Ｓ_／（ｘ、ｙ）、Ｓ_｜（ｘ、ｙ）、Ｓ_＼（ｘ、ｙ）から得られる特徴量は２値化後の各残存画素数に限られるものではなく例えば適当な閾値を越える最大輝度などを用いてもかまわない。
【００８０】
計算された特徴量を４次元ベクトルＵ＝（ｕ_１，ｕ_２，ｕ_３，ｕ_４）で表し、特徴量パターンと呼ぶ。今回の例では前述の￣での残存画素数ｕ_１、／での残存画素数ｕ_２｜での残存画素数ｕ_３、＼での残存画素数ｕ_４を特徴量とする。
【００８１】
更にそれぞれの残存画素数の比率から情報のエントロピーを計算することにより欠陥のテクスチャーの偏りを定量的に評価して新たな特徴量として扱ってもよい。エントロピーは事象の予測しやすさを表現する量であり事象ｌの生成確率をＰ_ｌ（ただし
【００８２】
【数７】

【００８３】
総和は起こる事象全てについてとる）とすると
【００８４】
【数８】

【００８５】
（総和は起こる全ての事象についてとる）
という式で表される。全ての事象が同じ確率で起こる最も予測しにくいとき情報のエントロピーは最大となり、ある事象だけが起こり（すなわちその生成確率１）他の事象が全く起こらない最も予測しやすいとき（すなわちその生成確率０）情報のエントロピーは最低値（０）になる。
【００８６】
ここではテクスチャーの方向性がランダムな場合は情報のエントロピーが大きく、テクスチャーの方向性が偏っている場合は情報のエントロピーが小さくなる。それぞれの残存画素数ｕ_１、ｕ_２、ｕ_３、ｕ_４が与えられたとき、
【００８７】
【数９】

【００８８】
のように
【００８９】
【数１０】

【００９０】
（総和は全てのカテゴリについてとる。すなわち、上記の場合、ｍ＝１〜４））となるよう規格化してから上式で情報のエントロピーｕ_５＝Ｈを計算して新たな特徴量としてもよい。このようにして、エントロピーを計算することにより、特定の方向性を持った欠陥か、方向性の無い欠陥かを判別することができる。
【００９１】
次に得られた各残存画素数から欠陥の種類および有害度について分類するための方法として、例えばベイズの定理に基づいた最尤決定を用いた方法の例を説明する。
【００９２】
まず、前述のように、欠陥について特徴量ｕ_１，ｕ_２，ｕ_３，ｕ_４からなるベクトルである特徴量パターンＵを求める。今回は４次元の特徴量パターンからＣ種類の欠陥種別又は欠陥程度を分類する
分類すべき欠陥種類または欠陥程度等をそれぞれカテゴリと呼ぶこととする。特徴量パターンＵ＝（ｕ_１，ｕ_２，ｕ_３，ｕ_４）が観測されたとき、その欠陥のカテゴリがｊである確率をＰ（ｊ｜Ｕ）、カテゴリｊ（カテゴリ数Ｃ）が特徴量パターンＵを生起する確率をＰ（Ｕ｜ｊ）、カテゴリｊの生起確率をＰ（ｊ）、特徴量パターンＵの生起確率をＰ（Ｕ）とすれば、ベイズの定理により
Ｐ（ｊ｜Ｕ）＝｛Ｐ（Ｕ｜ｊ）Ｐ（ｊ）｝／Ｐ（Ｕ）
【００９３】
ここでＰ（Ｕ｜ｊ）、Ｐ（ｊ）はあらかじめ多数のサンプルを集めて求めておくことができる。またＰ（Ｕ）は、通常はカテゴリによらないとする。そして、この式でｊを変えてＰ（ｊ｜Ｕ）が最大となるカテゴリｉを、欠陥のカテゴリ（欠陥種類または欠陥程度等）とすればよい。
【００９４】
実際には、Ｕ＝（ｕ_１，ｕ_２，ｕ_３，ｕ_４）は無限に存在するので、ｕ_１，ｕ_２，ｕ_３，ｕ_４のそれぞれについて、０から所定の最大値までの間を所定数Ｋに区分し、その区分毎にＵを決定する。このようにすれば、ＵはＫ^４個だけについて考慮すればよい。
【００９５】
あらかじめ、特徴量パターンＵとそのカテゴリ（欠陥種類または欠陥程度等）との対応の分かっている多量の学習用サンプルから教師付学習を行っておく。カテゴリｊの生成確率Ｐ（ｊ）を求め、また、カテゴリｊに分類された特徴量パターンＶ（Ｖ⊆Ｚ、Ｚは考えられるパターン全体の集合）のそれぞれの要素ｖ_ｋ（ｋ＝１〜４）のｋごとの平均値ｍ_ｋ、と分散σ_ｋｋ ^２、さらにｖ_ｋとｖ_ｌの共分散σ_ｋｌ ^２＝σ_ｌｋ ^２（ｋ≠ｌ）を計算しておく（ｋ＝１，２，３，４、ｌ＝１，２，３，４）。平均ベクトルをＭ_ｊ＝（ｍ_１，ｍ_２，ｍ_３，ｍ_４）^Ｔ，４×４の共分散行列を
【００９６】
【数１１】

【００９７】
と表す。ここで、
ｍ_１：学習用サンプルでの￣を構造要素としたときの残存画素数の平均
ｍ_２：学習用サンプルでの／を構造要素としたときの残存画素数の平均
ｍ_３：学習用サンプルでの｜を構造要素としたときの残存画素数の平均
ｍ_４：学習用サンプルでの＼を構造要素としたときの残存画素数の平均
σ_１１ ^２：学習用サンプルでの￣を構造要素としたときの残存画素数の分散
σ_２２ ^２：学習用サンプルでの／を構造要素としたときの残存画素数の分散
σ_３３ ^２：学習用サンプルでの｜を構造要素としたときの残存画素数の分散
σ_４４ ^２：学習用サンプルでの＼を構造要素としたときの残存画素数の分散
σ_１２ ^２＝σ_２１ ^２：学習用サンプルでの￣を構造要素としたときの残存画素数と／を構造要素としたときの残存画素数の共分散
σ_２３ ^２＝σ_３２ ^２：学習用サンプルでの／を構造要素としたときの残存画素数と｜を構造要素としたときの残存画素数の共分散
σ_３４ ^２＝σ_３４ ^２：学習用サンプルでの｜を構造要素としたときの残存画素数と＼を構造要素としたときの残存画素数の共分散
σ_１３ ^２＝σ_３１ ^２：学習用サンプルでの￣を構造要素としたときの残存画素数と｜を構造要素としたときの残存画素数の共分散
σ_２４ ^２＝σ_４２ ^２：学習用サンプルでの／を構造要素としたときの残存画素数と＼を構造要素としたときの残存画素数の共分散
σ_１４ ^２＝σ_４１ ^２：学習用サンプルでの￣を構造要素としたときの残存画素数と＼を構造要素としたときの残存画素数の共分散
である。
これらの平均ベクトルと分散行列は、全てのカテゴリｊについて計算しておく。
【００９８】
カテゴリｊを生起する確率が特徴量パターンＵのｎ次元（今回の場合はｎ＝４）正規分布に従うと仮定すると、ベイズの定理からの計算により、尤度関数は｜Σ_ｊ｜をΣ_ｊの行列式として
【００９９】
【数１２】

【０１００】
となる。
【０１０１】
ある検査で、特徴量パターンＷ（Ｗ⊆Ｚ）が検出されたとき、ｊを変えて、この尤度関数ｇ（ｊ｜Ｗ）が最大となるカテゴリｉを探せば、このｉでＰ（ｊ｜Ｗ）（特徴量パターンＷが観測されたときそれがカテゴリｊである確率）も最大となり、カテゴリｉに欠陥種類または欠陥程度を分類することができる。
【０１０２】
ここでは本発明の手法から得られる情報以外に欠陥信号の面積、輝度積算値、半値幅等の従来から使われてきた各種特徴量をも含めて認識してもよいので一般的にはｎ次元でのベイズの定理に基づいた最尤決定が考えられる。
【０１０３】
すなわち￣での残存画素数ｕ_１、／での残存画素数ｕ_２、｜での残存画素数ｕ_３、＼での残存画素数ｕ_４、情報のエントロピーｕ_５、欠陥信号の面積ｕ_６、輝度積算値ｕ_７、半値幅ｕ_８・・・を要素とするベクトルを特徴量パターンＵとして考える。各残存画素数からのパターン認識には上記したベイズの定理に基づいた最尤決定の他にツリー型の論理や最近傍決定則（ｎｅａｒｅｓｔ　ｎｅｉｇｈｂｏｒ　ｒｕｌｅ）あるいはニューラルネットワークの手法などさまざまなものが考えられる。
【０１０４】
以上の実施の形態は、特定形態欠陥の抽出または欠陥種類または欠陥程度の分類に主眼をおいて書かれているが、同様の原理から生産工程における処理むら等の無害な信号を欠陥候補から除外するために用いてもよい。
【０１０５】
以上の実施の形態は、単独の構造要素を用いたトップハット変換により、ノイズを除去する方法であった。この方法によると、画像の濃淡のように大きなうねりをもった大きなノイズを効果的に除去し、微小な欠陥を効果的に検出することができる。
【０１０６】
しかしながら、実際の表面欠陥を検出する場合には、微小ではあるが急峻な信号変化を持ったノイズが存在することがある。たとえば、受光素子であるＣＣＤの素子の感度むらに起因する画質の荒れなどがこのようなノイズに相当する。このような場合には、照明系に起因する緩やかな輝度斑と画質の荒れなどの急峻で細かいノイズ要素の両者を除去し、これらに対して中間的な急峻さの信号変化でを有する欠陥信号のみを抽出する必要がある。
【０１０７】
このような要望に適えるためには、サイズの異なる構造要素の対｛ｇ_＋（ｘ），　ｇ₋（ｘ）｝を使用し、それぞれのｃｌｏｓｉｎｇ演算またはｏｐｅｎｉｎｇ演算の差分を求めて中心輝度を通常１２８（即ち最大信号の半分の信号）にし、バンドパスフィルターにように中間的な信号変化をもつ欠陥のみを選択的に抽出すればよい。
【０１０８】
図１２に大きな半円状の構造要素ｇ_＋（ｘ）と小さな半円状の構造要素ｇ₋（ｘ）を対としてｃｌｏｓｉｎｇ演算を行い、その差分により欠陥検出を行う方法を１次元の例として説明する。
【０１０９】
図１２において（ａ）は元信号であり、図に示すように大きなうねりの中に、中程度のｘ軸方向の大きさと傾斜を有する欠陥と、ｘ軸方向の大きさは小さいが、急峻な傾斜を有するノイズとが混在している。
【０１１０】
これに対して、大きな半円状の構造要素ｇ_＋（ｘ）を用いてｃｌｏｓｉｎｇ演算を行うと、半円状の構造要素ｇ_＋（ｘ）は、欠陥信号の中にもノイズ信号に中にも入り込めず、（ｂ）に示すような軌跡をたどるので、結局、構造要素ｇ_＋（ｘ）によるｃｌｏｓｉｎｇ演算の結果は（ｃ）のようになる。
【０１１１】
また、小さな半円状の構造要素ｇ₋（ｘ）を用いてｃｌｏｓｉｎｇ演算を行うと、半円状の構造要素ｇ₋（ｘ）は、欠陥信号の中には入り込めるがノイズ信号の中には入り込めず、（ｄ）に示すような軌跡をたどるので、結局構造要素ｇ₋（ｘ）によるｃｌｏｓｉｎｇ演算の結果は（ｅ）のようになる。
【０１１２】
よって、（ｄ）と（ｅ）との差分をとるとその結果は（ｆ）のようになり、ノイズ信号と大きなうねりが除去され、欠陥信号が強調されて残る。図１２では模式的に１次元の負極性欠陥の例で説明したが、正極性の場合も、２次元の場合も、これから容易に想像できるとおりである。２次元の場合に使用する構造要素の対の例を図１３に示す。図１３において、数字は深さを示す。
【０１１３】
【実施例】
＜実施例１＞
以下、本発明を鉄鋼分野の鋼板の表面検査における有害欠陥抽出に適用した第１の例について述べる。
【０１１４】
画像処理装置内での信号処理の流れを図４に沿って説明する。先ずステップＳ１１において画像データを採取して入力する。次に、ステップＳ１２〜Ｓ１５において、それぞれ―、／、｜、＼方向に長い構造要素（例えば、図３の構造要素２のように長手方向に延びた楕円状の形態を有する構造要素）によりトップハット変換Ｓ_￣（ｘ、ｙ）、Ｓ_／（ｘ、ｙ）、Ｓ_｜（ｘ、ｙ）、Ｓ_＼（ｘ、ｙ）を行い、それぞれについてトップハット変換後の変換画像を得る。
【０１１５】
輝度の小さい欠陥を抽出したい場合は上からのトップハット変換、輝度の大きな欠陥を抽出したい場合は、下からのトップハット変換、両者を抽出したい場合は、両方のトップハット変換を行い、それぞれに変換後の変換画像を得る。上記のような構造要素によるトップハット変換により、構造要素の長手方向に急峻な信号変化をもつもののみを強調し、他の陰影をならした画像信号が得られる。
【０１１６】
次に、ステップＳ１６において、前述の計算によりＤ_ｍａｘ（ｘ、ｙ）とＤ_{ｌｉｎｅａｒ}（ｘ、ｙ）を求める。そして、ステップＳ１７においてＤ_ｍａｘ（ｘ、ｙ）とＤ_{ｌｉｎｅａｒ}（ｘ、ｙ）で形成されるグレースケール画像を出力し、ステップＳ１８でそれから２値化画像を得る。このときトップハット変換により注目する信号変化以外は慣らされ規格化することができるため適当な固定の閾値により、シェーディング補正を経た画像の場合と同様に容易に２値化を行うことができる。この段階で、人間が２値化画像を見て、欠陥を判別するようにしてもよい。
￣／｜＼それぞれの構造要素を使用してトップハット変換後に２値化した画像の例を図５に示す。
【０１１７】
また、方向性を有する構造要素を、０〜１８０°に亘って、４５°おきに回転させて、それぞれトップハット変換を行い、Ｄ_ｍａｘ（ｘ、ｙ）を求めた画像を図６に、Ｄ_{ｌｉｎｅａｒ}（ｘ、ｙ）を求めた画像を図７に示す。Ｄ_{ｌｉｎｅａｒ}（ｘ、ｙ）において、各構造要素のトップハット変換に対する重み係数は、横（−）方向の信号を出しやすくするため、Ａ₋＜Ａ_／＝Ａ_＼＜Ａ_｜のように調整してある。
【０１１８】
図６、図７のいずれにおいても、複雑な方向性を有する欠陥画像が、背景（ノイズ）に対して強調されて抽出されていることが分かる。
【０１１９】
この実施の形態では、その濃度積算値等の特徴量を求める。まず、ステップＳ１９で、２値化画像のうち、不要な孤立点を除去したうえで残存する画素を欠陥候補とする。次に、ステップＳ２０で、欠陥候補中の濃度（グレースケール値）の積算値、最低値（又は最高値）を求め、これらより欠陥の程度を判定する。また、欠陥候補と判定された２値化画像の面積より欠陥の大きさを求める。
【０１２０】
この実施の形態では、元画像にシェーディング補正をかけることは行っていない。シェーディング補正をかけると、その段階では欠陥の形状についての評価がおこわれてないため、鋼板上の大きな冷却水の流れなどを拾って過検出をおこしてしまう。またシェーディング処理においては、前述のように大きな正極性欠陥の信号のたちさがりの部分が反対の負極性の信号となり、正極性信号の影の部分を検出してしまうことになる。欠陥分類のための代表的なパラメータである濃度積算値は欠陥の領域で各画素の輝度を積算したものであるが、シェーディング処理によればこの特徴量を誤って計算してしまうことになる。
【０１２１】
よって、シェーディング補正を行わず、そこで元画像に対して欠陥より大きい適当な形の構造要素を使った下からのトップハット変換・上からのトップハット変換により負極性信号と正極性信号をそれぞれ抽出し、濃度積算値を求めれば適切な値が求められる。
【０１２２】
＜実施例２＞
以下、本発明を鉄鋼分野の鋼板の表面検査における有害欠陥抽出に適用した第２の例について述べる。
【０１２３】
画像処理装置内での信号処理の流れを図８に沿って説明する。先ずステップＳ２１において画像データを採取して入力する。次にステップＳ２２で、光学系に起因する輝度ムラを除去するためにシェーディング補正を行い、ステップＳ２３で２値化処理をおこなって欠陥候補を求める。そして、この欠陥候補のランレングスと中心輝度（２値化前）を求める。
【０１２４】
欠陥候補が挙げられると元画像に戻って、ステップＳ２６において、欠陥候補の近傍の画像を切り出す。これは、欠陥候補を含む画素と、その周辺の特定範囲の画素（例えば周囲の、ｘ、ｙ方向特定数の範囲の画素）を切り出すことによって求められる。そして、ステップＳ２８で、前述のようなグレースケールのモルフォロジーにおける、特定の形態を持つ構造要素を使用したトップハット変換の演算を行うが、本実施例においては、それに先立ち、ステップＳ２７において使用する構造要素を、ステップＳ２５で求められたランレングスと中心輝度とを使用して、図９に示すような表により求めている。
【０１２５】
すなわち、本実施例においては、ステップＳ２５で求められたランレングスと中心輝度を利用して予想される欠陥の形状を求め、その欠陥を有効に抽出できるような構造要素を、図９に示すような表から求めている。これにより、多種類の欠陥について、最適な構造要素でトップハット変換を行うことができる。
【０１２６】
ステップＳ２８でのトップハット変換で求められた画像について、ステップＳ２９で２値化処理を行い、２値化画像を得る。そして、ステップＳ３０で、２値化された残存画素数を算出し、ステップＳ３１で残存画素数が閾値以上であるかどうかを算出する。
【０１２７】
残存画素数が閾値以上の場合は、ステップＳ３２で欠陥として出力する。ステップＳ２６〜Ｓ３２は、抽出された欠陥候補の全てについて実施する。
【０１２８】
トップハット変換はシェーディング補正のかわりに元画像全体にわたっておこなってもよいが、処理速度を上げるため、本実施例では欠陥候補の近傍において行うこととする。このようにトップハット変換を施す領域を狭めても広い領域での移動平均が必要となるシェーディング補正と異なり、構造要素より充分に大きい面積であれば、その正規化の能力を弱めることはない。
【０１２９】
また、本手段においては、欠陥候補の特性に応じて構造要素を選択して適用しているので、元画像全体に亘って全ての構造要素を使用してトップハット変換を行う場合に比して、大幅に処理時間を短縮できる。
【０１３０】
＜実施例３＞
以下、本発明を鉄鋼分野の鋼板の表面検査における有害欠陥分類に適用した例について述べる。
【０１３１】
例えば冷間圧延の工程において例えば通常の欠陥が流れ方向と垂直につくのに対してロールの圧下による欠陥は鋼板の流れ方向につく傾向があり、鋼板の表面検査においてはこのようなロールマークを特定する必要がある。このように欠陥の分類において形態およびテクスチャーの評価は重要である。
【０１３２】
この処理の流れを図１０のフローチャートに従って説明する。まず先ずステップＳ４１において画像データを採取して入力する。次に、ステップＳ３２〜Ｓ５３において、￣／｜＼の方向性と特定のスケールを持つ複数の構造要素を使用してトップハット変換を行い、それぞれ、Ｓ_￣（ｘ、ｙ）、Ｓ_／（ｘ、ｙ）、Ｓ_｜（ｘ、ｙ）、Ｓ_＼（ｘ、ｙ）の画像を得て、適当な閾値により２値化することにより各残存画素数を算出し比較する。
【０１３３】
図１１に、構造要素の例を示す。各構造要素中の数字は、紙面に垂直な方向の深さに対応する。（ａ）はほぼ球状の構造要素であり、方向性を有しない欠陥の抽出に適している。（ｂ）はやや横長の楕円球状の構造要素であり、横長の欠陥を除外し、縦長の欠陥を検出するのに適している。（ｃ）は縦に長い角錐台状の構造要素であり、縦長の欠陥を除外し、横長の欠陥を検出するのに適している。個の実施例では、（ｃ）に示すような構造要素を使用した。｜の方向性をもった構造要素は、（ｃ）に対応するが、他の方向性をもった構造要素については、これを所定角度傾けたものとすればよい。￣／｜＼それぞれの構造要素を使用してトップハット変換後に２値化した画像の例を図５に示す。
【０１３４】
各残存画素数を比較して欠陥の種類および有害度について分類するためには、ステップＳ５４においてベイズの定理に基づいた最尤（ｍａｘｉｍｕｍ　ｌｉｋｅｌｉｈｏｏｄ）決定を行う。そして、その結果（欠陥分類結果）をステップＳ５５において出力する。
【０１３５】
＜実施例４＞
本発明を，鉄鋼分野の鋼板の表面検査における有害欠陥抽出に適用した第３の例について述べる。これは構造要素対を使用した例である。画像処理装置内での信号処理の流れを図１４に沿って説明する。先ず画像データを入力する（ステップＳ６１）、次に構造要素ｇ₋（ｘ，ｙ）を使用し、画像データに対してｃｌｏｓｉｎｇ演算を行い、その結果として演算画像Ｓ₋（ｘ，ｙ）を得る（ステップＳ６２）。
【０１３６】
続いて、構造要素ｇ₋（ｘ，ｙ）とはサイズの異なる構造要素ｇ_＋（ｘ，ｙ）を使用し、画像データに対してｃｌｏｓｉｎｇ演算を行、その結果として演算画像Ｓ_＋（ｘ，ｙ）を得る（ステップＳ６３）。そして、ステップＳ６４において演算画像Ｓ₋（ｘ，ｙ）とＳ_＋（ｘ，ｙ）の差分画像Ｄ（ｘ，ｙ）を求める。この際、２５６階調のグレースケールを有する差分画像に対して、１２８のバイアスをはかせ、差分画像のグレースケールが中央値近傍にくるようにする（ステップＳ６４）。
【０１３７】
次に、この画像信号を固定の閾値によって２値化し（ステップ６５）不要な孤立点を除去したうえで残存する画素を欠陥候補とする（ステップＳ６６）。このとき上記処理により注目する信号変化以外は慣らされ規格化することができるため適当な固定の閾値により、シェーディング補正を経た画像の場合と同様に容易に２値化を行うことができる。
【０１３８】
【発明の効果】
以上説明したように、本発明によれば、撮像によって得られた画像信号を新規な方法で処理することにより、欠陥の検出や種類の判別を精度よく行うことができる表面欠陥検出方法を提供することができる。
【図面の簡単な説明】
【図１】１次元関数ｆ（ｘ）に対して、半円状の構造関数ｇ１（ｘ）にてｏｐｅｎｉｎｇ演算を行った結果と、半円状の構造関数ｇ２（ｘ）にてｃｌｏｓｉｎｇ演算を行った結果を示す図である。
【図２】上からのトップハット変換について、１次元信号の場合を例にとって説明する図である。
【図３】トップハット変換に使用する構造要素の例と、トップハット変換の結果を示す図である。
【図４】本発明の第１の実施例における処理のフローを示すフローチャートである。
【図５】￣／｜＼それぞれの構造要素を使用してトップハット変換後に２値化した画像の例を示す図である。
【図６】方向性を有する構造要素を、０〜１８０°に亘って、４５°おきに回転させて、それぞれトップハット変換を行って求めた、Ｄ_ｍａｘ（ｘ、ｙ）を示す画像である。
【図７】方向性を有する構造要素を、０〜１８０°に亘って、４５°おきに回転させて、それぞれトップハット変換を行って求めた、Ｄ_{ｌｉｎｅａｒ}（ｘ、ｙ）を示す画像である。
【図８】本発明の第２の実施例における処理のフローを示すフローチャートである。
【図９】ランレングスと中心輝度からトップハット変換に使用する構造要素を決定する表の１例を示す図である。
【図１０】本発明の第３の実施例における処理のフローを示すフローチャートである。
【図１１】構造要素の例を示す図である。
【図１２】図１２に大きな半円状の構造要素ｇ_＋（ｘ）と小さな半円状の構造要素ｇ₋（ｘ）を対としてｃｌｏｓｉｎｇ演算を行い、その差分により欠陥検出を行う方法を説明する図である。
【図１３】２次元の場合に使用する構造要素の対の例を示す図である。
【図１４】構造要素の対を使用した処理における画像処理装置内での信号処理の流れを説明する図である。[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a technology for optically detecting a surface defect of a test object such as a film, paper, and steel.
[0002]
[Prior art]
As a method of optically detecting a defect generated on the surface of an industrial product, for example, as disclosed in Japanese Patent Application Laid-Open No. 8-82604, an image signal obtained by imaging is subjected to, for example, a defect width, a defect area, and a vertical and horizontal direction. There is known a method of classifying defects by a tree-type logic based on various characteristic amounts such as a ratio and a maximum intensity.
[0003]
Japanese Patent Application Laid-Open No. 8-189904 proposes a method of extracting a feature by causing a template designed in advance to correspond to a surface defect to act on image data or differential image data. Furthermore, the publication also proposes a method of generating a density co-occurrence matrix for image data or differential image data and obtaining a feature amount therefrom.
[0004]
[Problems to be solved by the invention]
However, all of the conventional techniques, including the techniques disclosed in the above-mentioned patent methods, are insufficient to accurately detect the surface defect of the inspection object.
[0005]
For example, a technique described in Japanese Patent Application Laid-Open No. 8-82604 is a means for obtaining an effective feature amount for a defect image form and its texture directionality, which are important factors for classifying a defect type or a harmfulness. The classification was incomplete without any. As for the method of knowing the directionality of the texture of an image, for example, a method of performing a two-dimensional Fourier transform on the image signal and performing spectrum analysis has been known, but the image signal does not have the same texture directionality over the entire area, An effective result could not be obtained when the defect image to be evaluated was localized.
[0006]
Further, in the technology described in Japanese Patent Application Laid-Open No. 8-189904, since the types of defects and the harmfulness to be judged are various, it is necessary to design at least 200 or more types of templates, and an optimum design is performed for each template. There is no clear guideline. Furthermore, in a method of generating a density co-occurrence matrix for image data or differential image data and calculating a feature amount from the density co-occurrence matrix, the image signal does not have the same texture directionality over the entire area, and a defect to be evaluated. Effective results could not be obtained when the image was localized. In addition, there is a problem that only a feature amount of a texture in a specific direction is extracted, and a geometric feature amount in which a more complicated form is used as a structural element cannot be extracted.
[0007]
In automatic defect detection, shading processing is used before binarization to normalize an image signal having uneven brightness. However, the falling part of the signal of a large positive defect is replaced with the opposite negative signal. In addition, a rising portion of the signal of a large negative polarity defect becomes a signal of the opposite positive polarity, and as a result, a non-defective signal is frequently erroneously detected as a defect.
[0008]
In this method, the moving average is calculated for a plurality of lines ahead and subtraction or division is performed on the target line, so that the falling portion of the positive polarity defect signal becomes a level lower than the center luminance after processing, and the negative polarity signal, the negative polarity defect This is because the rising portion of the signal has a level higher than the center luminance and becomes a signal of positive polarity. While it is important to determine the polarity of the defect signal in the automatic detection of the harmful defect, the differential processing such as a conventionally known shading method has the above-mentioned essential defects.
[0009]
Furthermore, regarding the classification of defects, defect classification is conventionally performed on image signals obtained by imaging using tree-type logic based on various characteristic amounts such as defect width, defect area, aspect ratio, and maximum intensity. . However, for example, regarding the form of the target defect image, whether it extends in the longitudinal direction, spreads in the width direction, spreads in the oblique direction, or has a more complicated luminance distribution, the harmfulness of the defect for that form is different.
[0010]
For example, there is no effective means when a defect signal having a steep change in the longitudinal direction and spreading in the width direction is particularly important, and there is a demand that only such a signal must be extracted.
[0011]
There is no means to extract only a specific form of the defect, which is an important factor for classifying the harmfulness of the defect, and overdetection often occurred. For example, many periodic defects caused by roll flaws appear as narrow and sharp signals in the longitudinal direction. Therefore, it is necessary to selectively extract and output such a form. However, there is no effective means for this.
[0012]
The present invention has been made in view of such circumstances, and provides a surface defect detection method that can accurately detect defects and determine the type by processing image signals obtained by imaging using a novel method. The task is to provide.
[0013]
[Means for Solving the Problems]
A first means for solving the above-mentioned problem is a step of obtaining a gray image by imaging a surface of an object to be inspected, and a method of obtaining a gray-scale morphology for at least a part of the obtained image signal. A surface defect detection method (claim 1) comprising a step of performing a top hat transform operation using a single or a plurality of structural elements having the following, and extracting a defect from an image obtained as a result. is there.
[0014]
As will be described later in detail in the embodiments of the present invention, when the top-hat transform is performed using a specific structural element in morphology, it is possible to remove a gradual change in an image and extract only a steep change. Therefore, it becomes easy to distinguish the surface defect from noise and extract it. Further, by changing the characteristics of the structural element, the type of the defect to be extracted can be changed. Furthermore, by performing top-hat conversion using a plurality of structural elements and comparing the converted images obtained as a result, the presence or absence of a defect can be more accurately detected.
[0015]
A second means for solving the above-mentioned problem is the first means, wherein “the step of extracting a defect from the resulting image” is performed by binarizing the resulting image, This is a step of extracting a defect based on the binarized image (claim 2).
[0016]
In the present means, an image obtained as a result of the top hat transformation is binarized with a predetermined threshold value, and a defect is extracted based on the obtained binarized image. Therefore, the defect can be detected by a simple method.
[0017]
A third means for solving the above-mentioned problem is the second means, wherein a top hat transform operation is performed using a plurality of structural elements. Is a step of extracting defects by calculating binary images corresponding to a plurality of obtained structural elements (claim 3). .
[0018]
In this means, a top hat transform operation is performed using a plurality of structural elements, the result is binarized, and a logical operation is performed on the binarized image corresponding to the obtained plurality of structural elements. , Defect extraction. Therefore, the presence or absence of a defect can be detected more accurately.
[0019]
A fourth means for solving the above-mentioned problem is the first means, wherein a top hat transform operation is performed using a plurality of structural elements, and a defect image is obtained from the resulting image. The step of extracting is a step of extracting a defect based on an image composed of a maximum value or a minimum value of pixel values obtained by each top hat transform (claim 4). is there.
[0020]
In the present means, a top hat transform operation is performed using a plurality of structural elements, and a defect is extracted based on an image consisting of the maximum or minimum pixel value obtained by each top hat transform.
[0021]
In this way, the defect is extracted based on the top hat transform using the structural element in which the defect is most easily detected, and the presence or absence of the defect can be accurately detected. For example, in the case where one defect has undulation, if an attempt is made to detect only one directional structural element, a part of the defect may be missed. A defect having a complicated shape can be accurately recognized as one defect.
[0022]
A fifth means for solving the above-mentioned problem is the first means, wherein a top hat transform operation is performed using a plurality of structural elements. The step of extracting is a step of extracting defects based on an image obtained by a linear combination of the obtained top hat transform results (claim 5).
[0023]
In this means, the defect is extracted based on the image obtained by the linear combination of the obtained top hat transform results, so that the defect of the target shape is emphasized by changing the coefficient of the linear combination. It is possible to accurately detect all the defects having a predetermined shape.
[0024]
A sixth means for solving the above-mentioned problems is a step of obtaining a gray image by imaging the surface of the object to be inspected, and a method of obtaining a gray-scale morphology for at least a part of the obtained image signal. Performs a top-hat transformation operation using a plurality of structural elements having the above, and performs at least one of defect classification and defect degree determination from the resulting image image of the top-hat transformation corresponding to the plurality of structural elements. A method for detecting a surface defect, comprising the step of:
[0025]
In this means, the type of the defect to be extracted can be changed by changing the characteristic of the structural element. Furthermore, by performing top-hat conversion using a plurality of structural elements and comparing the converted images obtained as a result, it is possible to accurately detect the type, degree, and type of the defect.
[0026]
A seventh means for solving the above-mentioned problem is the sixth means, wherein the image image after the top hat conversion is binarized, and the binarized image is used for at least classification of defects and determination of the degree of defects. The present invention is characterized in that the method includes a step of performing one of them.
[0027]
In the present means, an image obtained as a result of the top hat transformation is binarized with a predetermined threshold value, and a defect is extracted based on the obtained binarized image. Therefore, at least one of defect classification and defect degree determination can be performed by a simple method.
[0028]
Eighth means for solving the above-mentioned problems includes a step of obtaining a gray image by imaging the surface of the object to be inspected, and a directional property in a grayscale morphology for at least a part of the obtained image signal. Performs a top-hat transform operation using a plurality of structural elements having a specific shade, binarizes the resulting image image of the top-hat transform corresponding to the plurality of structural elements, and calculates remaining pixels. A method for calculating the entropy of information from these ratios and quantitatively evaluating the deviation in the directionality of the defect from the value of the entropy (claim 8).
[0029]
In this means, top hat transformation is performed using a plurality of structural elements having a specific directional shading, and the obtained image is binarized to calculate the remaining pixels. Then, entropy is calculated from the ratio. When the value of the entropy is large, it can be determined that the defect has less directivity, and when the value of the entropy is small, it can be determined that the defect is directional.
[0030]
A ninth means for solving the above-mentioned problem is any of the first means to the eighth means, wherein after a step of imaging a surface of the inspection object to obtain a gray image, another A method for determining a candidate portion of a defect and characteristics of the defect by the method, and performing a top-hat transform operation using a structural element corresponding to the characteristic of the defect for each candidate portion of the defect. (Claim 9).
[0031]
In this means, for example, the characteristics of the rough defect are obtained by an arbitrary conventional method, and the calculation of the top-hat transformation is performed using the structural element corresponding to the rough defect characteristic. It is possible to select and use various structural elements, and it is possible to reduce the operation time as compared with using all the structural elements. In addition, since a portion that is a candidate for a defect is determined in advance, the area for performing the top-hat conversion can be reduced.
[0032]
A tenth means for solving the above problem is the ninth means, wherein another method first performs shading processing and binarization on an image signal obtained by imaging to extract a defect candidate. Further, the present invention is characterized in that a part which is a candidate for a defect and a characteristic of the defect are obtained from the binarized image (claim 10).
[0033]
In this means, first, a shading process and binarization are performed on an image signal obtained by imaging to extract a defect candidate, and a defect candidate portion and the characteristic of the defect are obtained from the binarized image. Structural elements are used depending on the result. Therefore, structural elements can be properly used by simple preprocessing.
[0034]
An eleventh means for solving the above-mentioned problem is the tenth means, wherein a method for obtaining a candidate portion of a defect from the binarized image and a characteristic of the defect is performed by using a pixel in the binarized image. (Claim 11). The method according to claim 11, wherein the method is based on at least one of the center luminance and the run length before binarization.
[0035]
In this means, the binarized image indicates the approximate range of the defect. The central luminance of the pixel before binarization in that range represents the depth of the defect, and the run length represents the length of the defect. Therefore, by selecting a structural element based on at least one of them, an appropriate structural element can be selected based on at least one of the defect depth and the defect length.
[0036]
A twelfth means for solving the above-mentioned problems is a step of obtaining a gray image by imaging the surface of the object to be inspected, and a method of obtaining a gray-scale morphology for at least a part of the obtained image signal. A surface defect detection method comprising: performing a closing operation or an opening operation using a single or a plurality of structural element pairs having a difference to obtain a difference between the pairs; and extracting a defect from an image obtained as a result. (Claim 12).
[0037]
A thirteenth means for solving the above-mentioned problem is the twelfth means, wherein the step of extracting a defect from the obtained image is performed by binarizing the obtained image into a binary image. This is a step of extracting a defect based on the structured image (claim 13).
[0038]
A fourteenth means for solving the above-mentioned problem is the thirteenth means, wherein a difference between a closing operation or an opening operation is obtained by using a plurality of pairs of structural elements, and based on the binarized image. The step of extracting a defect is a step of extracting a defect by calculating a binary image corresponding to a plurality of obtained structural element pairs (claim 14). It is.
[0039]
A fifteenth means for solving the above-mentioned problem is the twelfth means, wherein a difference between a closing operation or an opening operation is obtained by using a plurality of pairs of structural elements, and an image obtained as a result is obtained. The step of extracting a defect from the image is a step of extracting a defect based on an image composed of a maximum value or a minimum value of pixel values obtained by a difference between each closing operation or opening operation. (Claim 15).
[0040]
A sixteenth means for solving the above-mentioned problem is the twelfth means, wherein a difference between operations of a closing operation or an opening operation is obtained by using a plurality of pairs of structural elements. The step of extracting a defect from the obtained image is a step of extracting a defect based on an image obtained by a linear combination of the obtained differences of each closing operation or opening operation. 16).
[0041]
A seventeenth means for solving the above-mentioned problems includes a step of obtaining a gray image by imaging the surface of the object to be inspected, and a method of obtaining a gray-scale morphology for at least a part of the obtained image signal. Classification of defects from a difference image of a closing operation or an opening operation corresponding to a plurality of structural element pairs obtained as a result of performing a difference operation of a closing operation or an opening operation using a plurality of structure element pairs having And a step of performing at least one of a defect degree determination and a surface defect detection method (claim 17).
[0042]
An eighteenth means for solving the above-mentioned problem is the seventeenth means, wherein the image image after the calculation of the difference of the closing operation or the opening operation is binarized, and the classification of the defect is performed from the binarized image. A surface defect detection method (claim 18), comprising a step of performing at least one of defect degree determination.
[0043]
A nineteenth means for solving the above problems is a step of obtaining a gray image by imaging the surface of the object to be inspected, and a method of obtaining a gray scale morphology for at least a part of the obtained image signal. A difference operation of a closing operation or an opening operation using a plurality of structural element pairs having a specific shading is performed, and the resulting difference image of the closing operation or the opening operation corresponding to the plurality of structural element pairs is represented by 2 A surface defect detection method (claim 19) comprising the steps of: calculating a residual pixel by calculating a value; calculating an entropy of information from these ratios; and quantitatively evaluating a bias of the directionality of the defect. is there.
[0044]
A twentieth means for solving the above-mentioned problem is any one of the twelfth means to the nineteenth means, wherein after a step of imaging a surface of the inspection object to obtain a gray image, another A method comprising: obtaining a candidate part of a defect and characteristics of the defect by a method; and performing a closing operation or an opening operation for each candidate part using a structural element pair corresponding to the characteristic of the defect. (Claim 20).
[0045]
A twenty-first means for solving the above-mentioned problem is the twentieth means, wherein another method first performs shading processing and binarization on an image signal obtained by imaging to extract a defect candidate. Further, the present invention is characterized in that a part which is a candidate for a defect and a characteristic of the defect are obtained from the binarized image (claim 21).
[0046]
A twenty-second means for solving the above-mentioned problem is the twenty-first means, wherein the method for obtaining a candidate portion of a defect from the binarized image and a characteristic of the defect is performed by using a method for determining a pixel in the binarized image. (Claim 22). The method according to claim 22, wherein the method is based on at least one of the center luminance and the run length before binarization.
[0047]
These twelfth to twenty-second means correspond to the first to twelfth means, respectively. In the first to twelfth means, the top-hat transform using a structural element is used, whereas in the thirteenth to twenty-fourth means, a pair of structural elements is used. The difference is that the difference of the result of the closing operation or the opening operation using the structural element of is calculated, the subsequent operation is performed based on the result of the difference operation, and the defect is detected based on the result. I have.
[0048]
By calculating the difference between the closing operation or the opening operation using the pair of structural elements, a defect having an intermediate size between the two structural elements can be emphasized and extracted. For example, if a large structural element and a small structural element are paired so that a large structural element can follow a gradual change such as luminance unevenness and a small structural element cannot follow a minute noise, Since both structural elements can follow the small noise and neither structural element can follow the minute noise, if a difference is obtained, these elements will not be detected. Then, it is possible to extract only a defect whose size is such that a large structural element cannot follow and a small structural element can follow.
[0049]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, embodiments of the present invention will be described with reference to the drawings. First, an image of the surface of the object to be inspected is stored in an image memory by an imaging system using a light source and a camera, for example. In the present embodiment, the obtained image signal is subjected to shading, the image is subjected to a top-hat transformation using a structuring element having a specific shading in a grayscale morphology, and a defect is determined from the resulting image. Is extracted.
[0050]
First, shading correction is performed on the image data collected in the image memory in order to remove luminance unevenness caused by the optical system. At this time, the normalization by shading correction is performed by taking a moving average in a predetermined range before and after the line of interest and subtracting or dividing from the signal of the line of interest, as is well known in the art. Since the image after the shading correction has a gray scale of, for example, 256 gradations, a binarization process can be performed by using a predetermined threshold value to obtain a binarized image.
[0051]
A defect candidate is obtained from the binarized image. That is, for example, an area where the pixel value is “1” in the binarized image is set as a defect candidate. The defect candidate includes a signal other than the defect signal such as a shadow portion which is a side effect of the shading correction.
[0052]
Next, a true defect signal is found from the defect candidates by a top-hat transformation known as an example of the morphological processing. That is, an image of a defect candidate and pixels in the vicinity thereof is cut out, and processing is performed using a specific structural element corresponding to the form of the defect signal of interest for top-hat transformation. The image signal used for this processing is a grayscale image before shading correction.
[0053]
The structural element used for the top-hat transformation may be one specific type. First, the binarized image is classified according to several patterns based on its length, aspect ratio, length direction, and the like. Other structural elements may be used. As an example of the structural element, for example, a plurality of structural elements having a direction of {/ |} and a specific scale may be used for simplicity. The direction of each line represents a structural element having a long shape in that direction. Each structural element is not limited to this, and may adopt a more complicated form.
[0054]
A plurality of processed images S obtained after the top hat transform for each structural element_￣(X, y), S_/(X, y), S_|(X, y), S_＼Maximum value of (x, y)
D_max(X, y) = max ｛S_￣(X, y), S_/(X, y), S_|(X, y), S_＼(X, y)｝
, A linear defect long in a direction perpendicular to each direction can be extracted as a center.
[0055]
Also, the coefficient A for each pixel of the processed image_￣, A_/, A_＼, A_|Weighted linear combination
D_linear(X, y)
= A_￣・ S_￣(X, y) + A_/・ S_/(X, y) + A_|・ S_|(X, y) + A_＼・ S_＼(X, y)
May be taken. At this time, for example, A_￣<A_/= A_＼<A_|In particular, a horizontal linear defect becomes particularly sensitive, and a vertical linear defect can be weighted. Further, it is also possible to binarize the obtained plurality of images with a predetermined threshold value and execute an operation such as a logical sum or a logical product to extract a true harmful defect.
[0056]
The morphology of the gray scale is described in detail in, for example, “Morphology” (Corona). Morphology is a new image processing technology that is being applied to the analysis of medical images, such as breast cancer X-ray screening, and uses a set-theoretic operation that uses preset structural elements to extract a specific geometric structure from an original image. This is to selectively extract only the elements that have.
[0057]
The morphological processing on the grayscale image is based on an element called a structural element that moves the image and an operation called Minkowski sum and Minkowski difference. Typical operations include a dilation operation and an erosion operation, and an opening operation and a closing operation of a combination thereof. There is an operation.
[0058]
When the gray value of the two-dimensional image to be processed is expressed as a function f (X) in a domain F of a two-dimensional space, and the gray value of a structural element is expressed as a structure function g (X) in a domain G of a two-dimensional space (X is 2 Dimensional variable, U appearing below is also a two-dimensional variable), and in the case of discrete data, first the Minkowski sum is
[0059]
(Equation 1)

[0060]
The Minkowski difference is
[0061]
(Equation 2)

[0062]
Is defined as Note that, as a symbol for defining the Minkowski difference, a symbol in which-is inserted in a circle is usually used. However, such a symbol cannot be used due to font restrictions. Further, the dilation operation is performed by converting g (X) of the Minkowski sum to a symmetric g^s(X) = g (−X)
[0063]
(Equation 3)

[0064]
In addition, the erosion (scraping) operation is g (X) of the Minkowski sum to g^s(X)
[0065]
(Equation 4)

[0066]
Is defined by The g (U) is translated in parallel with the function f (X), and the maximum value of the sum of the corresponding points in the vertical direction is a dilation operation, and the minimum value of the difference is an erosion operation.
Here, the opening operation is the combination of the Minkowski sum and the erosion operation.
[0067]
(Equation 5)

[0068]
The closing operation combines the Minkowski difference and the dilation operation.
[0069]
(Equation 6)

[0070]
It becomes.
[0071]
FIG. 1 shows a result obtained by performing an opening operation on a one-dimensional function f (x) using a semicircular structure function g1 (x) and a result obtained by performing a closing operation using a semicircular structure function g2 (x). The results obtained are shown. In this case, the opening operation corresponds to a smoothing process from below, and the closing operation corresponds to a smoothing process from above.
[0072]
If the value of the pixel subjected to the smoothing process from the bottom by the opening operation is subtracted from the value of the original pixel, the value of the pixel subjected to the top-hat transformation from the bottom and the value of the pixel subjected to the smoothing process from the top by the closing operation is Subtraction from the value of the original pixel results in top-hat transformation from above. From the viewpoint of extracting a defect signal from an image, top-hat conversion from below means extraction of a positive defect signal, and top-hat conversion from above means extraction of a negative defect signal.
[0073]
The top-hat transform from above will be described with reference to FIG. 2 using a one-dimensional signal as an example. Where the value of the function changes slowly, the sphere can trace it exactly, but if there is a negative pulse, the sphere will not be able to enter that part, so it will be out of the traceable range It will be off. Therefore, a smooth curve is obtained as the trajectory of the sphere, as shown in FIG. If the trajectory of the sphere is subtracted from the original signal, only a sharp negative polarity signal can be extracted (c).
[0074]
As apparent from FIG. 2, what kind of signal to extract is determined by the form of the structural element. Also in the two-dimensional case, it can be easily imagined that only the defect signal having a sharp change can be extracted by using the structural element as shown in FIG.
[0075]
In FIG. 3, as shown in FIG. 3A, a top-hat transform is performed on a structural element having a shape of a downwardly convex elliptical sphere. (B) is an image before conversion (a gray image, which is a binarized image due to restrictions in application documents), and has a sharp gradation change among gentle gradation changes. There are two defects. The actual size of the structural element is the size shown on the lower side of (b). (C) shows the gray image after the top hat conversion, and it can be seen that only the defect is extracted.
[0076]
Furthermore, when classifying the types and harmfulness of defects as well as extracting true harmful defects, it is conceivable to perform top-hat conversion with a plurality of structural elements. In the following example, for the sake of simplicity, a case will be described in which a plurality of structural elements having a direction of {/ |} and a specific scale are used. However, the present invention is not limited to this, and a more complicated form may be used as a structural element. You may adopt it.
[0077]
As described above, the top hat transformation is performed using the structural element, and the detection, the degree determination, and the type determination of the defect are performed based on the image after the top hat transformation. The method of detecting the defect, determining the degree, and determining the type based on the image is not particularly limited, and a conventionally used method may be appropriately used, or in an extreme case, the image after the top hat conversion may be used by a human. You may look and do these things.
[0078]
Hereinafter, an example of defect type determination will be described. First, a method of determining whether a defect has directivity or non-directionality will be described.
[0079]
First, after top-hat transformation in each structural element, S_￣(X, y), S_/(X, y), S_|(X, y), S_＼(X, y) is obtained, a plurality of binarizations are performed using an appropriate threshold, and the number of remaining pixels is calculated. In this example, four structural elements having the directionality of {/ |} and a specific scale are used to obtain the number of remaining pixels of each of the four images as shown in FIG. At this time, a plurality of processed images S_￣(X, y), S_/(X, y), S_|(X, y), S_＼The feature amount obtained from (x, y) is not limited to the number of remaining pixels after binarization, and for example, a maximum luminance exceeding an appropriate threshold may be used.
[0080]
The calculated feature amount is converted into a four-dimensional vector U = (u₁, U₂, U₃, U₄) And is called a feature amount pattern. In this example, the number of remaining pixels u₁, / Remaining number of pixels u₂| The number of remaining pixels u₃, 残存 the number of remaining pixels u₄Is a feature amount.
[0081]
Further, the bias of the texture of the defect may be quantitatively evaluated by calculating the entropy of information from the ratio of the number of remaining pixels, and may be treated as a new feature amount. Entropy is a quantity expressing the predictability of an event, and the generation probability of an event l is represented by P_l(However,
[0082]
(Equation 7)

[0083]
The sum is taken for all events that occur)
[0084]
(Equation 8)

[0085]
(Sum is taken for all events that occur)
It is expressed by the formula. The entropy of information is greatest when all events occur with the same probability and is the least predictable, and when only one event occurs (ie its generation probability 1) and when no other events occur most easily (ie its generation probability 0) ) The entropy of the information becomes the lowest value (0).
[0086]
Here, when the directionality of the texture is random, the entropy of the information is large, and when the directionality of the texture is uneven, the entropy of the information is small. Each remaining pixel number u₁, U₂, U₃, U₄Is given,
[0087]
(Equation 9)

[0088]
like
[0089]
(Equation 10)

[0090]
(The sum is obtained for all categories; that is, in the above case, m = 1 to 4)), and then the information entropy u₅= H may be calculated as a new feature value. By calculating the entropy in this manner, it is possible to determine whether the defect has a specific direction or not.
[0091]
Next, an example of a method using maximum likelihood determination based on Bayes' theorem will be described as a method for classifying the type and harmfulness of a defect from the obtained numbers of remaining pixels.
[0092]
First, as described above, the feature amount u₁, U₂, U₃, U₄A feature amount pattern U, which is a vector composed of In this case, the C defect types or defect degrees are classified from the four-dimensional feature pattern.
The type of defect or the degree of defect to be classified is called a category. Feature amount pattern U = (u₁, U₂, U₃, U₄) Is observed, P (j | U) is the probability that the category of the defect is j, P (U | j) is the probability that category j (the number of categories C) will generate a feature pattern U, and category j Is P (j) and the probability of occurrence of the feature pattern U is P (U), the Bayes theorem gives
P (j | U) = {P (U | j) P (j)} / P (U)
[0093]
Here, P (U | j) and P (j) can be obtained by collecting a large number of samples in advance. It is assumed that P (U) does not normally depend on the category. Then, by changing j in this equation, the category i that maximizes P (j | U) may be set as the defect category (defect type or defect degree, etc.).
[0094]
In practice, U = (u₁, U₂, U₃, U₄) Is infinite, so u₁, U₂, U₃, U₄Are divided into a predetermined number K from 0 to a predetermined maximum value, and U is determined for each of the divisions. In this way, U becomes K⁴You only need to consider the individual.
[0095]
In advance, supervised learning is performed from a large number of learning samples in which the correspondence between the feature amount pattern U and its category (defect type or defect degree, etc.) is known. The generation probability P (j) of the category j is obtained, and each element v of the feature amount pattern V (V⊆Z, Z is a set of all possible patterns) classified into the category j is obtained._kAverage value m for each k of (k = 1 to 4)_k, And variance σ_kk ²And v_kAnd v_lCovariance σ of_kl ²= Σ_lk ²(K ≠ l) is calculated (k = 1, 2, 3, 4; l = 1, 2, 3, 4). Average vector is M_j= (M₁, M₂, M₃, M₄)^T, 4 × 4 covariance matrix
[0096]
[Equation 11]

[0097]
It expresses. here,
m₁: Average number of remaining pixels when ￣ in the learning sample is a structural element
m₂: Average number of remaining pixels when / is used as a structural element in the learning sample
m₃: Average number of remaining pixels when | is a structural element in the learning sample
m₄: Average number of remaining pixels when ＼ in the learning sample is a structural element
σ₁₁ ²: Variance of the number of remaining pixels when ￣ in the training sample is a structural element
σ₂₂ ²: Variance of the number of remaining pixels when / is used as a structural element in the learning sample
σ₃₃ ²: Variance of the number of remaining pixels when | is a structural element in the learning sample
σ₄₄ ²: Variance of the number of remaining pixels when ＼ in the training sample is a structural element
σ₁₂ ²= Σ₂₁ ²: Covariance of the number of remaining pixels when ￣ is a structural element and the number of remaining pixels when / is a structural element in the learning sample
σ₂₃ ²= Σ₃₂ ²: Covariance between the number of remaining pixels in the learning sample when / is a structural element and the number of remaining pixels when | is a structural element
σ₃₄ ²= Σ₃₄ ²: Covariance between the number of remaining pixels when | is a structural element and the number of remaining pixels when ＼ is a structural element in the learning sample
σ₁₃ ²= Σ₃₁ ²: The covariance of the number of remaining pixels when ￣ is a structural element and the number of remaining pixels when | is a structural element in the learning sample
σ₂₄ ²= Σ₄₂ ²: Covariance of the number of remaining pixels in the learning sample when / is a structural element and the number of remaining pixels when と き is a structural element
σ₁₄ ²= Σ₄₁ ²: The covariance of the number of remaining pixels when と し た is a structural element and the number of remaining pixels when ＼ is a structural element in the learning sample
It is.
The average vector and the variance matrix are calculated for all the categories j.
[0098]
Assuming that the probability of occurrence of the category j follows the n-dimensional (n = 4 in this case) normal distribution of the feature pattern U, the likelihood function becomes | Σ by calculation from Bayes' theorem._j|_jAs the determinant of
[0099]
(Equation 12)

[0100]
It becomes.
[0101]
When a feature amount pattern W (W⊆Z) is detected in a certain inspection, if j is changed to find a category i in which the likelihood function g (j | W) is maximized, P (j | W) (the probability that the feature amount pattern W is category j when it is observed) is also maximum, and the defect type or defect degree can be classified into category i.
[0102]
Here, in addition to the information obtained from the method of the present invention, various types of conventionally used features such as the area of the defect signal, the integrated luminance value, and the half width may be recognized. The maximum likelihood decision based on Bayes' theorem in is considered.
[0103]
That is, the number of remaining pixels u in ￣₁, / Remaining number of pixels u₂, | The number of remaining pixels u₃, 残存 the number of remaining pixels u₄, Entropy of information u₅, The area u of the defect signal₆, Integrated luminance value u₇, Half width u₈.. Are considered as feature amount patterns U. For the pattern recognition from each number of remaining pixels, various methods such as a tree-type logic, a nearest neighbor rule (nearest @ neighbor @ rule), or a neural network method can be considered in addition to the maximum likelihood determination based on Bayes' theorem described above. .
[0104]
Although the above embodiment has been described with an emphasis on the extraction of specific form defects or the classification of defect types or defects, harmless signals such as uneven processing in the production process are excluded from defect candidates based on the same principle. May be used to do so.
[0105]
The above-described embodiment is a method of removing noise by top-hat transformation using a single structural element. According to this method, it is possible to effectively remove a large noise having a large undulation such as a shading of an image, and to effectively detect a minute defect.
[0106]
However, when an actual surface defect is detected, there is a case where noise having a small but sharp signal change exists. For example, rough image quality due to uneven sensitivity of a CCD element serving as a light receiving element corresponds to such noise. In such a case, both a gentle luminance unevenness due to the illumination system and a steep and fine noise element such as a rough image quality are removed, and a defect signal having an intermediate steep signal change with respect to these is removed. Only need to extract.
[0107]
In order to meet such a demand, a pair of structural elements having different sizes is required.₊(X), g₋(X) Using 輝 度, find the difference between each closing operation or opening operation to set the center luminance to normally 128 (ie, half the maximum signal), and to have a defect with an intermediate signal change like a band pass filter. Only what is necessary is to extract only.
[0108]
FIG. 12 shows a large semicircular structural element g.₊(X) and small semicircular structural element g₋A method of performing a closing operation using (x) as a pair and detecting a defect based on the difference will be described as a one-dimensional example.
[0109]
In FIG. 12, (a) is an original signal, as shown in the figure, a defect having a medium size and inclination in the x-axis direction and a small size in the x-axis direction, Noise having a slope is mixed.
[0110]
On the other hand, a large semicircular structural element g₊When a closing operation is performed using (x), a semicircular structural element g₊(X) cannot enter into the defect signal nor the noise signal, and follows the trajectory as shown in (b).₊The result of the closing operation by (x) is as shown in (c).
[0111]
Also, a small semicircular structural element g₋When a closing operation is performed using (x), a semicircular structural element g₋(X) can enter into the defect signal but cannot enter into the noise signal, and follows the trajectory shown in (d).₋The result of the closing operation by (x) is as shown in (e).
[0112]
Therefore, when the difference between (d) and (e) is obtained, the result is as shown in (f), the noise signal and the large swell are removed, and the defective signal remains emphasized. Although FIG. 12 schematically illustrates an example of a one-dimensional negative defect, both positive and two-dimensional defects can be easily imagined from now on. FIG. 13 shows an example of a pair of structural elements used in the two-dimensional case. In FIG. 13, the numbers indicate the depth.
[0113]
【Example】
<Example 1>
Hereinafter, a first example in which the present invention is applied to the extraction of harmful defects in the surface inspection of a steel sheet in the steel field will be described.
[0114]
The flow of signal processing in the image processing device will be described with reference to FIG. First, in step S11, image data is collected and input. Next, in steps S12 to S15, each of the structural elements is elongated in the −, /, |, and ＼ directions (for example, a structural element having an elliptical shape extending in the longitudinal direction like the structural element 2 in FIG. 3). Hat transformation S_￣(X, y), S_/(X, y), S_|(X, y), S_＼(X, y) is performed to obtain a converted image after the top hat conversion for each.
[0115]
If you want to extract defects with low brightness, perform top-hat conversion from the top.If you want to extract defects with high brightness, perform top-hat conversion from the bottom.If you want to extract both, perform top-hat conversion for both. Obtain a converted image after conversion. By the above-described top-hat transform using the structural element, only an image signal having a steep signal change in the longitudinal direction of the structural element is emphasized, and an image signal in which another shading is smoothed is obtained.
[0116]
Next, in step S16, D_max(X, y) and D_linear(X, y) is obtained. Then, in step S17, D_max(X, y) and D_linearA grayscale image formed by (x, y) is output, and a binary image is obtained therefrom in step S18. At this time, except for the signal change of interest by the top-hat transformation, it can be used and standardized, so that binarization can be easily performed with an appropriate fixed threshold value in the same manner as in the case of an image that has undergone shading correction. At this stage, a human may look at the binarized image and determine the defect.
FIG. 5 shows an example of an image binarized after the top hat transform using each of the structural elements {/ |}.
[0117]
Further, the directional structural elements are rotated every 45 ° from 0 ° to 180 ° to perform a top-hat transformation, and D_maxThe image for which (x, y) was determined is shown in FIG._linearFIG. 7 shows an image obtained for (x, y). D_linearIn (x, y), the weighting factor for the top hat transform of each structural element is A₋<A_/= A_＼<A_|It has been adjusted as follows.
[0118]
6 and 7, it can be seen that a defect image having a complicated directionality is extracted with emphasis on the background (noise).
[0119]
In this embodiment, a characteristic amount such as an integrated value of the density is obtained. First, in step S19, pixels remaining after removing unnecessary isolated points from the binarized image are determined as defect candidates. Next, in step S20, the integrated value and the minimum value (or the maximum value) of the density (gray scale value) in the defect candidate are obtained, and the degree of the defect is determined based on these values. Further, the size of the defect is obtained from the area of the binarized image determined as the defect candidate.
[0120]
In this embodiment, shading correction is not performed on the original image. When the shading correction is performed, since the shape of the defect is not evaluated at that stage, a large flow of cooling water on the steel plate is picked up and overdetection is performed. In the shading process, as described above, the portion of the signal of the large positive polarity defect becomes the opposite negative polarity signal, and the shadow portion of the positive polarity signal is detected. The integrated density value, which is a typical parameter for defect classification, is obtained by integrating the luminance of each pixel in the defect area. However, according to the shading processing, this feature amount is erroneously calculated.
[0121]
Therefore, the shading correction is not performed, and the negative and positive signals are respectively extracted by top-hat conversion from the bottom and top-hat conversion from the top using a structural element of an appropriate shape larger than the defect in the original image. Then, if an integrated value of the concentration is obtained, an appropriate value can be obtained.
[0122]
<Example 2>
Hereinafter, a second example in which the present invention is applied to the extraction of harmful defects in the surface inspection of a steel sheet in the steel field will be described.
[0123]
The flow of signal processing in the image processing apparatus will be described with reference to FIG. First, in step S21, image data is collected and input. Next, in step S22, shading correction is performed to remove luminance unevenness caused by the optical system, and in step S23, binarization processing is performed to obtain defect candidates. Then, the run length and center luminance (before binarization) of the defect candidate are obtained.
[0124]
If a defect candidate is given, the process returns to the original image, and in step S26, an image near the defect candidate is cut out. This is obtained by cutting out a pixel including a defect candidate and a pixel in a specific range around the pixel (for example, surrounding pixels in a specific number range in the x and y directions). Then, in step S28, a top-hat transformation operation using a structural element having a specific form in the above-described gray scale morphology is performed. In the present embodiment, prior to this, the structure used in step S27 is used. Elements are obtained from a table as shown in FIG. 9 using the run length and the center luminance obtained in step S25.
[0125]
That is, in the present embodiment, as shown in FIG. 9, structural elements that can be used to determine the shape of an expected defect by using the run length and the center luminance obtained in step S25 and effectively extract the defect are shown in FIG. From a simple table. As a result, the top hat transformation can be performed on the various types of defects using the optimal structural elements.
[0126]
The image obtained by the top-hat transformation in step S28 is subjected to a binarization process in step S29 to obtain a binarized image. Then, in step S30, the binarized number of remaining pixels is calculated, and in step S31, it is calculated whether the number of remaining pixels is equal to or greater than a threshold.
[0127]
If the number of remaining pixels is equal to or larger than the threshold value, it is output as a defect in step S32. Steps S26 to S32 are performed for all of the extracted defect candidates.
[0128]
The top-hat conversion may be performed over the entire original image instead of the shading correction. However, in order to increase the processing speed, in this embodiment, the top-hat conversion is performed near the defect candidate. Unlike shading correction, which requires a moving average over a wide area even when the area to be subjected to the top hat transform is narrowed in this way, if the area is sufficiently larger than the structural element, the normalization ability is not reduced.
[0129]
Further, in the present means, since the structural element is selected and applied according to the characteristics of the defect candidate, it is compared with the case where the top hat transform is performed using all the structural elements over the entire original image. , Can greatly reduce the processing time.
[0130]
<Example 3>
Hereinafter, an example in which the present invention is applied to the classification of harmful defects in the surface inspection of a steel sheet in the steel field will be described.
[0131]
For example, in the process of cold rolling, for example, normal defects are attached perpendicular to the flow direction, whereas defects caused by roll reduction tend to be attached in the flow direction of the steel sheet. Need to be identified. Thus, evaluation of morphology and texture is important in defect classification.
[0132]
The flow of this processing will be described with reference to the flowchart of FIG. First, in step S41, image data is collected and input. Next, in steps S32 to S53, a top-hat transformation is performed using a plurality of structural elements having the directionality of {/ |} and a specific scale._￣(X, y), S_/(X, y), S_|(X, y), S_＼An image of (x, y) is obtained and binarized by an appropriate threshold to calculate and compare the number of remaining pixels.
[0133]
FIG. 11 shows an example of a structural element. The number in each structural element corresponds to the depth in the direction perpendicular to the plane of the paper. (A) is a substantially spherical structural element, which is suitable for extracting a defect having no directivity. (B) is a slightly horizontally long elliptical spherical structural element, which is suitable for detecting a vertically long defect while excluding a horizontally long defect. (C) is a vertically long truncated pyramid-shaped structural element, which is suitable for detecting a horizontally long defect while excluding a vertically long defect. In the examples, a structural element as shown in (c) was used. A structural element having a direction of | corresponds to (c), but a structural element having another direction may be inclined at a predetermined angle. FIG. 5 shows an example of an image binarized after the top hat transform using each of the structural elements {/ |}.
[0134]
In order to compare the numbers of remaining pixels and classify the types and harmfulness of defects, a maximum likelihood (maximum likelihood) determination is performed in step S54 based on Bayes' theorem. Then, the result (defect classification result) is output in step S55.
[0135]
<Example 4>
A third example in which the present invention is applied to the extraction of harmful defects in the surface inspection of a steel sheet in the steel field will be described. This is an example using a structural element pair. The flow of signal processing in the image processing apparatus will be described with reference to FIG. First, image data is input (step S61), and then the structural element g₋Using (x, y), a closing operation is performed on the image data, and as a result, the operation image S₋(X, y) is obtained (step S62).
[0136]
Then, the structural element g₋Structural element g different in size from (x, y)₊Using (x, y), a closing operation is performed on the image data, and as a result, the operation image S₊(X, y) is obtained (step S63). Then, in step S64, the calculated image S₋(X, y) and S₊A difference image D (x, y) of (x, y) is obtained. At this time, a bias of 128 is applied to the difference image having the gray scale of 256 gradations, so that the gray scale of the difference image is close to the median (step S64).
[0137]
Next, the image signal is binarized using a fixed threshold (step 65), and unnecessary pixels are removed, and the remaining pixels are determined as defect candidates (step S66). At this time, except for the signal change of interest by the above processing, it can be used and standardized, so that binarization can be easily performed with an appropriate fixed threshold value in the same manner as in the case of an image that has undergone shading correction.
[0138]
【The invention's effect】
As described above, according to the present invention, there is provided a surface defect detection method capable of accurately detecting a defect and determining a type by processing an image signal obtained by imaging by a novel method. be able to.
[Brief description of the drawings]
FIG. 1 shows a result obtained by performing an opening operation on a one-dimensional function f (x) using a semicircular structure function g1 (x) and a result obtained by performing a closing operation using a semicircular structure function g2 (x). It is a figure showing the result of having performed.
FIG. 2 is a diagram illustrating top-hat transformation from above, taking a one-dimensional signal as an example;
FIG. 3 is a diagram illustrating an example of a structural element used for a top-hat transformation and a result of the top-hat transformation.
FIG. 4 is a flowchart illustrating a processing flow according to the first embodiment of the present invention.
FIG. 5 is a diagram showing an example of an image binarized after top-hat transformation using each of the structural elements {/ |}.
FIG. 6 is a graph showing a D obtained by rotating a structural element having directionality every 45 ° from 0 ° to 180 °, and performing a top-hat transformation, respectively._maxIt is an image showing (x, y).
FIG. 7 is a diagram showing a D obtained by rotating a structural element having directionality every 45 ° from 0 ° to 180 ° and performing a top-hat transformation, respectively._linearIt is an image which shows (x, y).
FIG. 8 is a flowchart illustrating a flow of processing according to a second embodiment of the present invention.
FIG. 9 is a diagram showing an example of a table for determining a structural element used for a top hat transform from a run length and a center luminance.
FIG. 10 is a flowchart illustrating a flow of processing according to a third embodiment of the present invention.
FIG. 11 is a diagram showing an example of a structural element.
FIG. 12 shows a large semicircular structural element g₊(X) and small semicircular structural element g₋FIG. 9 is a diagram illustrating a method of performing a closing operation with (x) as a pair and performing defect detection based on the difference.
FIG. 13 is a diagram showing an example of a pair of structural elements used in a two-dimensional case.
FIG. 14 is a diagram illustrating a flow of signal processing in the image processing apparatus in processing using a pair of structural elements.

Claims (22)

Imaging a surface of an object to be inspected to obtain a gray image, and using one or more structural elements having a specific shading in a gray scale morphology for at least a part of the obtained image signal. A method for detecting a surface defect, comprising a step of performing a Hat transformation operation and extracting a defect from an image obtained as a result. 2. The surface defect detection method according to claim 1, wherein the step of extracting a defect from the image obtained as a result includes binarizing the image obtained as a result, and extracting the defect based on the binarized image. A method for detecting a surface defect. 3. The surface defect detection method according to claim 2, wherein a top hat transform operation is performed using a plurality of structural elements, and a step of extracting a defect based on the binarized image is obtained. A defect extraction step by calculating a binarized image corresponding to the plurality of structural elements. 2. The surface defect detection method according to claim 1, wherein a top hat transform operation is performed using a plurality of structural elements, and the step of extracting a defect from an image obtained as a result includes: A surface defect detection method, which is a step of extracting a defect based on an image composed of a maximum value or a minimum value of pixel values obtained by the hat transformation. The surface defect detection method according to claim 1, wherein a top hat transform operation is performed using a plurality of structural elements, and a step of extracting a defect from a resulting image is obtained. A defect extraction process based on an image obtained by a linear combination of the respective top-hat transformation results. Imaging a surface of an object to be inspected to obtain a gray image, and a top hat transform using a plurality of structural elements having a specific shading in a grayscale morphology for at least a part of the obtained image signal. Performing a calculation, and performing at least one of defect classification and defect degree determination from the image image after the top hat conversion corresponding to the plurality of structural elements obtained as a result, . 7. The surface defect detection method according to claim 6, further comprising a step of binarizing the image image after the top hat conversion, and performing at least one of defect classification and defect degree determination from the binarized image. Characteristic surface defect detection method. Imaging a surface of an object to be inspected to obtain a gray image, and using a plurality of structural elements having a specific directional shading in a grayscale morphology for at least a part of the obtained image signal. Performs a top-hat transform operation, binarizes the resulting top-hat transformed image corresponding to a plurality of structural elements, calculates remaining pixels, calculates the entropy of information from these ratios, A method for detecting a surface defect, comprising a step of quantitatively evaluating a deviation in directionality of a surface. The surface defect detection method according to any one of claims 1 to 8, wherein after a step of imaging a surface of the inspection object to obtain a gray image, a defect candidate is identified by another method. A surface defect detection method comprising: obtaining a characteristic portion and a characteristic of the defect; and performing a top-hat conversion operation for each candidate portion of the defect using a structural element corresponding to the characteristic of the defect. 10. The surface defect detection method according to claim 9, wherein another method first performs shading processing and binarization on an image signal obtained by imaging to extract a defect candidate, and detects a defect from the binarized image. A surface defect detection method for determining a candidate portion of a target and characteristics of the defect. The surface defect detection method according to claim 10, wherein the method of obtaining a candidate for a defect from the binarized image and a characteristic of the defect includes determining a center of a pixel in the binarized image before binarization. A method for detecting a surface defect, wherein the method is a method for obtaining the defect based on at least one of luminance and run length. Imaging a surface of an object to be inspected to obtain a gray image, and using one or more pairs of structural elements having a specific shading in a grayscale morphology for at least a part of the obtained image signal. A surface defect detection method comprising a step of performing a closing operation or an opening operation to obtain a difference between the two, and extracting a defect from an image obtained as a result. 13. The surface defect detection method according to claim 12, wherein the step of extracting a defect from the image obtained as a result includes binarizing the image obtained as a result, and extracting the defect based on the binarized image. A method for detecting a surface defect. 14. The surface defect detection method according to claim 13, wherein a difference between a closing operation or an opening operation is obtained using a plurality of pairs of structural elements, and a step of extracting a defect based on the binarized image is included. A method for extracting defects by calculating binary images corresponding to a plurality of obtained structural element pairs. The surface defect detection method according to claim 12, wherein a difference between a closing operation or an opening operation is obtained using a plurality of pairs of structural elements, and a step of extracting a defect from an image obtained as a result is included. A step of extracting a defect based on an image including a maximum value or a minimum value of pixel values obtained by a difference between each closing operation or opening operation. 13. The surface defect detection method according to claim 12, wherein a difference between a closing operation or an opening operation is obtained using a plurality of pairs of structural elements, and a step of extracting a defect from an image obtained as a result is included. A step of extracting a defect based on an image obtained by a linear combination of the obtained differences of the closing operation or the opening operation. Capturing a surface of the object to be inspected to obtain a gray image; and performing a closing operation or a closing operation using a plurality of pairs of structural elements having a specific shading in a grayscale morphology on at least a part of the obtained image signal. a step of performing a difference operation of the opening operation and performing at least one of classification of a defect and determination of a degree of the defect from an image image of the difference of the closing operation or the opening operation corresponding to the plurality of structural element pairs obtained as a result A method for detecting a surface defect, comprising: The surface defect detection method according to claim 17, wherein the image image after the difference calculation of the closing operation or the opening operation is binarized, and at least one of defect classification and defect degree determination is performed from the binarized image. A method for detecting a surface defect, comprising the step of: Imaging a surface of an object to be inspected to obtain a gray image, and using a plurality of pairs of structural elements having a specific directional shading in a grayscale morphology for at least a part of the obtained image signal. The difference between the closing operation or the opening operation is calculated, and the resulting difference image of the closing operation or the opening operation corresponding to the plurality of structural element pairs is binarized to calculate the remaining pixels. A method for detecting a surface defect, comprising a step of calculating entropy of information and quantitatively evaluating a deviation in the directionality of a defect. The surface defect detection method according to any one of claims 12 to 19, wherein after a step of imaging a surface of the inspection target object to obtain a gray image, a defect candidate is determined by another method. And calculating the difference between the closing operation and the opening operation for each candidate part of the defect by using a structural element pair corresponding to the characteristic of the defect. Surface defect detection method. 21. The surface defect detection method according to claim 20, wherein another method first performs shading processing and binarization on an image signal obtained by imaging to extract a defect candidate, and detects a defect from the binarized image. A surface defect detection method for determining a candidate portion of a target and characteristics of the defect. 22. The surface defect detection method according to claim 21, wherein the method of obtaining a candidate portion of a defect from the binarized image and a characteristic of the defect is performed by: A method for detecting a surface defect, wherein the method is a method for obtaining the defect based on at least one of luminance and run length.

Images