JP2000242782A - Apparatus and method for detecting object in image and recording medium recording this method - Google Patents

Abstract

(57) [Summary] To detect an object from an image, when a recognition model is created by image processing such as KL conversion on a partial image of a local rectangular area including the object, the orientation of the object is determined. It cannot be detected when there is an apparent change such as occlusion. SOLUTION: A plurality of images including an object to be detected and an object other than the object are input (101, 102), and a color, an edge, a texture, or a feature vector by a combination thereof is extracted from each input image (111). The space of the feature vector is hierarchically classified into an image characterization region of the object and other regions (112), and a recognition model is created in this feature space (113). In addition, the local region is detected by selecting a local region to be collated by scanning a search window with respect to an unknown image, extracting a feature vector from each local region, and comparing the extracted feature vector with the recognition model. It is determined whether or not the detection target object is included in the unknown image by the matching calculation.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a detection device and a detection method for detecting an object in an image, and more particularly to a method for creating a recognition model of a detection target object from a plurality of images in which the detection target object is captured. The present invention relates to an apparatus and a method for determining whether or not a detection target object is included in an unknown image based on the recognition model, and detecting a local region in which the object is included when the object is included.

[0002]

2. Description of the Related Art In recent years, many techniques have been proposed for an apparatus for creating a recognition model for recognizing and detecting an object in an image. One of them is an apparatus for directly creating a recognition model of an object to be detected from a plurality of images of the object.

[0003] This conventional apparatus uses a KL (Kalfunen) for a case set consisting of an image of an object to be detected or an image of an object to be detected, which is a partial image of a local rectangular area including the object. (Reve) A two-dimensional template image of the object to be detected is created by applying an image processing method such as transformation, and a recognition model of the object to be detected is created by the template image itself. The object is detected based on the collation calculation between many local regions of the template image and this template image. For example, there is the following document.

Reference 1: M. Turk and A.
Pentland, "Eigenfaces for
Recognition ", J. Cognitive
Neuroscience, Vol. 3, No. 1, pp. 7
1-86 (1991) In this reference 1, in order to recognize and detect a human face from an image, a KL transform is applied to a set of face images photographed from the front of a plurality of humans, so that an unknown image can be obtained. A face image recognition model capable of detecting a face image facing forward is created.

[0005]

SUMMARY OF THE INVENTION In order to detect an object to be detected from an image, an image of the object to be detected or a partial image of a local rectangular area including the object of the image of the object to be detected. In a conventional apparatus that creates a two-dimensional template image of a target object as a recognition model by applying an image processing method such as KL transformation to a set of cases such as In the matching calculation, since the similarity of each local region of the unknown image and the template image with respect to the entire image is measured as the matching evaluation in the matching calculation, even when the detection target object is included in the local region, the detection is performed. If the target object has any apparent change such as orientation or occlusion, it cannot be detected because the similarity of the whole image is not evaluated. There is a problem.

The problem to be solved by the present invention is that when an object to be detected is photographed in an unknown image, the object can be detected even when there is a slight change in orientation or occlusion. An object of the present invention is to propose a device and a method for creating a possible recognition model and detecting a detection target object based on the recognition model.

[0007]

According to the present invention, there is provided an apparatus for detecting an object in an image, comprising the steps of: receiving an input object based on a recognition model created from a plurality of images in which an object to be detected is captured; It is determined whether or not the detection target object is included in the unknown image, and if so, a device for detecting an object in an image that detects a local region in which the object is captured,
Means for inputting a plurality of images in which the object to be detected is captured; means for inputting a plurality of images in which objects other than the object to be detected are captured; and Means for extracting a color, an edge, a texture, or a feature vector by a combination thereof from each input image other than the object, and the extracted detection target object and the feature vector other than the detection target object, thereby extracting a feature space of the detection target object. The image processing apparatus further includes means for hierarchically classifying the image feature into a characteristic area and other areas, and means for creating a recognition model of the detection target object in the classified feature space.

The local area detecting means scans a search window of the input unknown image over the entire image to select a local area to be collated, and the local area detecting means selects each of the selected local areas. Means for extracting a feature vector from, and determining whether or not the detection target object is present in the unknown image by a collation calculation between each of the extracted feature vectors and the created recognition model of the detection target object. And a means for detecting the local area when it exists.

According to another aspect of the present invention, there is provided a method for detecting an object in an image, comprising the steps of: detecting an object in an input unknown image based on a recognition model created from a plurality of images of the object to be detected; It is a method of detecting an object in an image to determine whether or not a detection target object is included, and if so, a local region in which the object is detected. A plurality of images in which the target object is captured, a plurality of images in which the objects other than the detection target object are captured are input, and the colors, edges, textures, or each of the input images other than the detection object and the detection object is input. By extracting a feature vector by combining these, the extracted detection target object and a feature vector other than the detection target object are used to convert a feature space into a feature region of the image feature of the detection target object and the other Hierarchically classified into frequency, characterized by creating a recognized model of the detection target object in the classified feature space.

In the detection of the local region, a search window is scanned over the entirety of the input unknown image to select a local region to be collated, and a feature vector is selected from each of the selected local regions. Is extracted, and it is determined whether or not the detection target object is included in the unknown image by a collation calculation between the extracted feature vectors and the created recognition model of the detection target object. It is characterized by detecting a local region.

A program for causing a computer to execute the processing procedure in the detection method is recorded on a recording medium readable by the computer.

[0012]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 shows the configuration of an apparatus for creating a recognition model according to the present invention. Means 101, 1 of this device configuration
02, 111 to 113 are a mouse, an image scanner,
It is realized by using a so-called computer system including a display, a CPU, a memory device, and the like.

First, the image input means 101 uses an image scanner or the like to input a plurality of images in which an object to be detected is captured, and stores the plurality of images in a memory or the like. Here, if the detection target object is included in a local rectangular area of the input image, the rectangular area is designated with a mouse or the like, a new partial image is created, and this partial image is stored in a memory or the like. To be stored.

Next, the image input means 102 inputs a plurality of images of objects other than the object to be detected, and stores the images in a memory or the like.

Next, the means 111 to 113 execute the following processing on the CPU to automatically create a recognition model of the detection target object in the processing system.

First, feature vector extraction processing means 111
Is a color from each input image in which the object to be detected input by the image input means 101 and 102 and the others are captured,
Extract feature vectors by edges, textures, or their combination.

Next, the hierarchical classification processing means 112 classifies the feature space hierarchically based on the set of the extracted feature vectors of the object to be detected and the object vector other than the object to be detected. Must not include the elements of both sets at the same time.

Finally, recognition model creation processing means 113
Generates the union of each local region in the feature space containing the feature vector of the detection target object, and generates the union set as the characterization region of the image feature of the detection target object, thereby generating the recognition model of the detection target object. create.

Next, a specific example of processing relating to the creation of a recognition model according to the present invention will be described with reference to FIGS.

First, a plurality of images of the object to be detected are input by the image input means 101. However, when the detection target object is included in a certain local rectangular area in the input image, a circumscribed rectangular area is specified so that the detection target object is included and the background is not included as much as possible, and a new partial image is created. Then, this partial image is input.

Next, a plurality of images, other than the object to be detected, are input by the image input means 102.

Next, the feature vector extraction processing means 111 extracts a color, an edge, a texture, or a feature vector based on a combination thereof from the input object to be detected and each image other than the object to be detected.

As for the feature vector extraction, for example, “Y-I Ohta, T. Kanda and T.
Sasaki, “Color information
for region segmentation ”,
Comp. and Img. Proc. , 13: 222-241
(1980) "(Ref. 2) and" Yoichi Mizuno, Aki Kobayashi, Toshiyuki Yoshida, Yoshinori Sakai, Image Search by Edge Direction Feature,
IEICE Technical Report, IE 98-62 (1998) ”(reference 3).

For example, as a color feature vector, HSI
A color histogram obtained by dividing the color system hue into 16 (the number of dimensions is 1)
6) can be adopted. As the edge feature vector, an edge direction histogram having four directions (4 dimensions) can be adopted. This means that, for each pixel of the image, there are four horizontal, vertical, 45 °, 135 °
Find the absolute value of the Gaussian derivative in the direction,
When the maximum value of these four absolute values satisfies a certain threshold, the pixel is determined as an edge point, and the direction in which the maximum value is given is determined as the edge direction of the pixel. Next, a histogram is created by counting the number of corresponding edge points in the image in each of the horizontal, vertical, 45 °, and 135 ° directions. Each component of the color and edge histograms is normalized by dividing by the total number of pixels of the image.

Next, the hierarchical classification processing means 112
And a set Sn of feature vectors extracted from each image other than the detection target object, which is a set Spos of feature vectors extracted from each image of the detection target object.
The feature space V is hierarchically classified according to the following method using eg, and the Spos element and the Sneg element are not simultaneously included in each local region of the characteristic space V formed by the classification. Hereinafter, a region in the feature space V where the image feature of the detection target object is characterized is referred to as a positive region, and the other region is referred to as a negative region.

First, the division processing means 201 divides each axis of the feature space V into two equal parts, and ^{sets the} feature space V to 2 ^dimV (dimV is the number of dimensions of the feature space V, dimV =
20) Divide into square areas.

Next, the positive / negative judgment processing means 202
As shown in (1), for each rectangular area in the feature space V formed by the division processing means 201, it is determined whether it is a positive area or a negative area in the following manner. This determination is as follows: (a) If the rectangular area includes only the elements of the set Spos, the area is determined as a positive area.

(B) If the rectangular area includes only the elements of the set Sneg, the area is defined as a negative area. Set Spo
A negative area is also defined when neither element of s nor Sneg is included.

Finally, the end determination processing means 203 ends the processing if all the rectangular areas of the feature space V are classified as positive or negative. Square W ₁ not classified positive and negative, ..., if there is Wn (including elements of the set Spos and Sneg simultaneously) are rectangular regions W _1, ..., W
n, the division processing means 201 and the positive / negative judgment processing means 20
2, the processing is performed, and the rectangular areas W ₁ ,..., Wn are subdivided until the area is classified as either positive or negative.

Finally, in the recognition model creation processing means 113, a union area is generated for each local area including the elements of the set Spos, and is generated as a characterization area of the image feature of the detection target object. Then, a recognition model of the detection target object is created.

FIG. 4 is a block diagram of the object detecting apparatus of the present invention based on the recognition model created as described above.
This apparatus configuration is actually realized on a computer system.

First, the unknown image input means 301 inputs an unknown image from a camera, an image recording medium or the like. The search window scan processing unit 311 scans a search window of various scales over the entire image with respect to the unknown image input from the input unit 301 and selects a large number of local rectangular regions for matching.

Next, feature vector extraction processing means 312
Extracts a feature vector from each local rectangular area selected by the search window scan processing means 311.

Finally, the recognition model matching processing means 313
By the collation calculation of each extracted feature vector and the recognition model created and stored using the above-described recognition model creation device, it is determined whether or not the detection target object is included in the unknown image, If it is, the area is detected.

Next, a specific processing example of the apparatus for detecting an object from an unknown image based on the recognition model of the present invention will be described with reference to FIG.

The unknown image X is input, and the search window scan processing means 311 scans a rectangular search window of various scales over the entire image as shown in FIG. .

Next, the feature vector extraction processing means 312 extracts a feature vector from each selected local rectangular area. However, the feature vector extracted from each local rectangular area in the unknown image is the same as the feature vector used for creating the recognition model of the detection target object.

Finally, the recognition model matching processing means 313 determines whether or not the unknown image X includes the detection target object by a comparison calculation with the recognition model of the detection target object. Detect the area. That is, if at least one feature vector extracted from the selected local rectangular area in the unknown image X is included in the R area of the feature space, that is, the characterization area of the image feature of the detection target object, this unknown image It is determined that X includes the detection target object, and the object is detected as being included in the local rectangular area from which the feature vector included in the R area is extracted.

In the present invention, some or all of the apparatuses shown in FIGS. 1, 2 and 4 can be made to function using a computer. In addition, the means of each drawing is described in a computer program so that it can be executed, and the computer program can be executed by a computer-readable recording medium such as an FD.
(A floppy disk), an MO, a ROM, a memory card, a CD, a DVD, a removable disk, and the like.

[0041]

As described above, according to the present invention,
By providing a plurality of images of the object to be detected and objects other than the object to be detected, the processing system learns the image features related to the color, edge, texture, or a combination thereof of the object to be detected and detects the object to be detected. Automatically create a recognition model for Since image features such as colors, edges, and textures are local image features of the object, stable and robust extraction is achieved in actual processing. For this reason, even in the process of detecting an object from an unknown image, even when the object to be detected is captured in the unknown image and there is a slight change in the orientation, occlusion, etc. When a local rectangular area including the target object and substantially no background is selected as a target to be compared, a desired image feature can be extracted from the area. Recognition and identification of the object in the leverage region can be performed. Therefore, it is possible to detect the area of the detection target object from the unknown image.

[Brief description of the drawings]

FIG. 1 is a processing configuration diagram according to a recognition model creation of the present invention.

FIG. 2 is a processing configuration diagram for classifying a feature space into a feature area of an image feature of a detection target object and other areas.

FIG. 3 is an example in which a feature space is classified into a feature area of an image feature of a detection target object and other areas.

FIG. 4 is a processing configuration diagram relating to detection of a target object in an image according to the present invention.

FIG. 5 is a specific example of processing for determining whether or not a target object is included in an unknown image and detecting a local rectangular area when the target object is included.

[Explanation of symbols]

 101: Image input means showing an object to be detected 102: Image input means showing an object other than the object to be detected 111 ... Feature vector extraction processing means 112 ... Hierarchical classification processing means 113 ... Recognition model creation processing means 201 ... Division processing Means 202 ... Positive / negative judgment processing means 203 ... End judgment processing means 301 ... Unknown image input means 311 ... Search window scan processing means 312 ... Feature vector extraction processing means 313 ... Recognition model collation processing means

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Toshiaki Sugimura 3-19-2 Nishishinjuku, Shinjuku-ku, Tokyo F-term (reference) within Nippon Telegraph and Telephone Corporation 5L096 BA11 DA02 FA19 FA35 FA41 GA17 GA19 GA51 JA09 JA11 KA13 MA07 9A001 HH21 HH23 KK31

Claims (5)

[Claims] 1. A method for determining whether or not a detection target object is included in an input unknown image based on a recognition model created from a plurality of images in which the detection target object is captured. An apparatus for detecting an object in an image that detects a local region in which the object is captured, wherein the recognition model creating unit includes: a unit that inputs a plurality of images in which the detection target object is captured; and the detection target object. Means for inputting a plurality of images in which the other objects have been captured, means for extracting a color, an edge, a texture, or a feature vector by combining these from the input image other than the detection target object and the detection target object; and Means for hierarchically classifying the feature space into a characteristic region of the image feature of the detection target object and a region other than the region based on the detection target object and the feature vector other than the detection target object. Means for creating a recognition model of a detection target object in the classified feature space. An apparatus for detecting an object in an image, comprising: 2. A method for determining whether or not a detection target object is included in an input unknown image based on a recognition model created from a plurality of images in which the detection target object is captured. An apparatus for detecting an object in an image for detecting a local region in which the object is captured, wherein the local region detecting unit scans a search window for the input unknown image over the entire image and performs matching with the matching target. Means for selecting a local region, and means for extracting a feature vector from each of the selected local regions. An unknown image is obtained by a collation calculation between each of the extracted feature vectors and the created recognition model of the detection target object. A means for determining whether an object to be detected is present in the image and detecting a local area when the object is present in the image. 3. It is determined whether or not a detection target object is included in an input unknown image based on a recognition model created from a plurality of images in which the detection target object is captured. A method for detecting an object in an image that detects a local region in which the object is captured, wherein the generation of the recognition model includes inputting a plurality of images in which the detection target object is captured, and capturing images other than the detection target object. A plurality of extracted images, and extracting a color, an edge, a texture, or a feature vector by a combination thereof from each of the input images other than the detection target object and the detection target object, and extracting the extracted detection target object and the detection target. The feature space is hierarchically classified into a feature area of the image feature of the detection target object and other areas by the feature vector other than the object, and the detection object is detected in the classified feature space. A method for detecting an object in an image, wherein a recognition model of the object is created. 4. It is determined whether or not a detection target object is included in an input unknown image based on a recognition model created from a plurality of images in which the detection target object is captured. A method for detecting an object in an image that detects a local region in which the object is captured, wherein the local region is detected by scanning a search window of the input unknown image over the entire image. A local region is selected, a feature vector is extracted from each of the selected local regions, and a detection target object is detected in an unknown image by a collation calculation between the extracted feature vectors and the created recognition model of the detection target object. A method for detecting an object in an image, comprising determining whether or not the image is captured and detecting a local region when the image is captured. 5. An image, wherein a program for causing a computer to execute a processing procedure in the method for detecting an object in an image according to claim 3 or 4 is recorded on a recording medium readable by the computer. A recording medium that records a method of detecting an object in the medium.