[go: up one dir, main page]

CN103186790A - Object detecting system and object detecting method - Google Patents

Object detecting system and object detecting method Download PDF

Info

Publication number
CN103186790A
CN103186790A CN2011104567794A CN201110456779A CN103186790A CN 103186790 A CN103186790 A CN 103186790A CN 2011104567794 A CN2011104567794 A CN 2011104567794A CN 201110456779 A CN201110456779 A CN 201110456779A CN 103186790 A CN103186790 A CN 103186790A
Authority
CN
China
Prior art keywords
window
window area
confidence
image
area
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN2011104567794A
Other languages
Chinese (zh)
Inventor
王强
毛文涛
马赓宇
金智渊
金培亭
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Samsung Telecommunications Technology Research Co Ltd
Samsung Electronics Co Ltd
Original Assignee
Beijing Samsung Telecommunications Technology Research Co Ltd
Samsung Electronics Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Samsung Telecommunications Technology Research Co Ltd, Samsung Electronics Co Ltd filed Critical Beijing Samsung Telecommunications Technology Research Co Ltd
Priority to CN2011104567794A priority Critical patent/CN103186790A/en
Publication of CN103186790A publication Critical patent/CN103186790A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Image Analysis (AREA)

Abstract

本发明提供一种对象检测系统和方法。所述对象检测系统包括：图像接收单元，接收待检测图像；特征提取单元，利用根部模板对待检测图像进行扫描，以提取多个窗口区域的图像特征；可变形部件模型检测单元，通过将提取的多个窗口区域的图像特征输入可变形部件模型，以利用可变形部件模型获得所述多个窗口区域的置信度，其中，可变形部件模型调整每个窗口区域的大小使得每个窗口区域的置信度达到最大；对象确定单元，根据窗口区域的置信度确定存在对象的窗口区域。

Figure 201110456779

The present invention provides an object detection system and method. The object detection system includes: an image receiving unit, which receives an image to be detected; a feature extraction unit, which scans the image to be detected by using a root template to extract image features of a plurality of window regions; a deformable part model detection unit, which extracts The image features of multiple window regions are input into the deformable part model to obtain the confidence of the multiple window regions by using the deformable part model, wherein the deformable part model adjusts the size of each window region so that the confidence of each window region The degree reaches the maximum; the object determination unit determines the window area where the object exists according to the confidence degree of the window area.

Figure 201110456779

Description

Object detection systems and method
The technology neighborhood
The present invention relates to vision and area of pattern recognition.More particularly, relate to a kind of object detection systems and method based on deformable component model (DPM, Deformable Part Model).
Background technology
Object detection is an important technology in the vision technique, and it has very important application in intelligent video surveillance, content-based image/video retrieval, image/video note, auxiliary man-machine interaction.Because different classes of object has a great difference in shape, so object detection is very difficult.
Multiple object detection scheme has been proposed so far, for example, Boosting method, DPM etc.Boosting method use characteristic is trained a plurality of simple Weak Classifiers, then these Weak Classifiers is configured to cascade classifier, in image each sliding window is classified.Yet the Boosting method can successfully detect comparatively simple object such as people's face, human eye, but still can not satisfy the detection of common object (for example, automobile etc.).DPM comes parametrization DPM by the outward appearance of each parts in the image and for the geometric model that obtains the spatial relationship between the parts.The study of DPM parameter can be expressed as the classification problem of using hidden variable (latent variable), and can use latent support vector machine (SVM) to solve this classification problem.DPM has represented the current development level in this field, wins the championship title in the PASCAL VOC in 2009 about object detection.DPM is very effective with respect to additive method, can handle piece image several seconds for a classification.Yet for the occasion that real-time is had relatively high expectations, such speed still can not meet the demands.In addition, still there is the problem that is difficult to the overlapping a plurality of object instances in the detected image in DPM.
DPM generally includes: the data item of the object root (root) of image/parts (part); Measure the deformatter of the distortion cost of these parts from the anchor station of each parts.Object instance can be expressed as followsin in the score of DPM:
f ( p 0 , . . . p n ) = Σ i = 0 n F i · φ ( H , p i ) - Σ i = 1 n d i · φ d ( dx i , dy i ) + b - - - ( 1 )
Wherein,
Figure BSA00000648239000012
Be the data item in the parts definition of the root of object and object,
Figure BSA00000648239000013
It is deformatter.
Here, p 0The root of indicated object, p 1, p 2... p nThe n of an indicated object parts, the quantity of the parts of n indicated object, it is positive integer, F iBe that (i equals to represent the convolution filter corresponding with the root proper vector at 0 o'clock for the convolution filter corresponding with root proper vector and component feature vector, i is not equal at 0 o'clock and represents the convolution filter corresponding with the component feature vector), H is the characteristics of image pyramid of input picture, φ (H, p i) expression by the characteristics of image pyramid obtain at p iFeature, φ d(dx i, dy i)=(dx i, dy i, (dx i) 2, (dy i) 2), dx i, dy iRepresent i parts in the horizontal direction with vertical direction on skew, d iBe the parameter of deformatter, b is the skew of the equation (1) as scoring function, and it depends on the concrete DPM of use.
In addition, also can be represented as based on the score in the sorter of DPM:
f(z)＝β·Ψ(H，z)， (2)
Wherein, z=(p 0... p n), β represents the model parameter of DPM; Z is latent vector, has comprised the numbering of position, scaling ratio and/or the concrete root template of using of object root and parts.
When equation 1 and equation 2 are consistent,
β＝(F 0，...F n，d 1，...d n，b)；
Ψ(H，z)＝(φ(H，p 0)，...φ(H，p n)，-φ d(dx 1，dy 1)，...-φ d(dx n，dy n)，1)。
In above-mentioned model, similar to traditional SVM, can obtain β to this model training by utilizing positive negative sample, also, parameter F i, d iAnd b.
In the process of the object in detected image, use root template (that is, detection window) that image is scanned to extract a plurality of window areas (that is, root), with the input as DPM of the characteristics of image of window area.The window that must assign to determine to exist object according to each window area.
Yet, in traditional DPM, at each root template or component model, use be that the length and width of fixing are recently mated object, therefore having among a small circle with the template length breadth ratio, the object of difference is easy to be lost in detection.
Summary of the invention
One object of the present invention is to solve technical matters above-mentioned.
An aspect of of the present present invention provides a kind of object detection systems, comprising: image receiving unit receives image to be detected; Feature extraction unit is utilized the root template to treat detected image and is scanned, to extract the characteristics of image of a plurality of window areas; Deformable component model detecting unit, characteristics of image input deformable component model by a plurality of window areas that will extract, to utilize the deformable component model to obtain the degree of confidence of described a plurality of window areas, wherein, the deformable component model size of adjusting each window area makes the degree of confidence of each window area reach maximum; The object determining unit is determined to exist the window area of object according to the degree of confidence of window area.
Alternatively, adjust the scope of each window area that extracts, the corresponding convolution filter of each window area after feasible the adjustment reaches maximum with the dot product of the characteristics of image on each window area after the adjustment, makes the degree of confidence of each window area reach maximum.
Alternatively, the deformable component model obtains by training, wherein, when training deformable component model, adjusts the size as the window area of sample, makes that the degree of confidence as the window area of sample reaches maximum.
Alternatively, when training deformable component model, adjustment is as the scope of the window area of sample, the corresponding convolution filter of the window area as sample after make adjusting with reach maximum at the dot product as the characteristics of image on the window area of sample after the adjustment, make that the degree of confidence as the window area of sample reaches maximum.
Alternatively, described deformable component model is to mix the deformable component model.
Alternatively, described object detection systems also comprises: the redundant unit that suppresses, from the described a plurality of window areas that obtained degree of confidence, remove pseudo-window area according to the interactive relation between described a plurality of window areas, wherein, the redundant unit that suppresses comprises: the characteristic information extraction unit, from each window area characteristic information extraction; The redundant unit of removing utilizes the characteristic information that extracts to determine described interactive relation, to remove pseudo-window area from described a plurality of window areas.
Alternatively, described characteristic information comprises at least one in the degree of confidence of degree of confidence, parts of the positional information of the positional information of degree of confidence, root of window area and/or yardstick information, parts and/or yardstick information, root.
Alternatively, the redundant unit of removing is by maximizing following equation and judge and remove pseudo-window area:
Figure BSA00000648239000031
Wherein, M represents the quantity of described a plurality of window areas;
φ (x i, y i)=y iX i
Figure BSA00000648239000032
x i=(v i(s), Z), v i(s) degree of confidence of i window area of expression, Z represents the vector of K dimension, K represents the quantity of the deformable component model that described mixing deformable component model is included, the v of Z i(c) individual element is that other elements of 1, Z are zero, v i(c) expression detects the index of i the employed deformable component model of window area; y iThat represents i window area is used for whether sign is the binary score of pseudo-window area; y iWhether what expression was used for j window area is the binary score of pseudo-window area for sign;
Figure BSA00000648239000033
The representation model parameter, d IjRepresent the interactive relation between i window area and j the window area,
Wherein, when described equation maximized, the window area with binary score of the pseudo-window area of sign was judged as pseudo-window area.
Alternatively, use the φ (x of precognition i, y i),
Figure BSA00000648239000041
SS divides class methods by predetermined structure to be trained to obtain
Figure BSA00000648239000042
Alternatively, the interactive relation between the window area comprise that root-root is mutual, root-parts mutual, at least one in mutual of parts-parts.
Alternatively, described object detection systems also comprises: context deformable component model detecting unit, the contextual feature input context sorter of the window area of degree of confidence will have been obtained, to obtain the new degree of confidence of window area, wherein, the context sorter is to utilize as the contextual feature of sample to train the sorter that obtains.
Alternatively, contextual feature comprises: shaped position feature, neighborhood characteristics, the collaborative feature that occurs.
Alternatively, the size of shaped position character representation window area in image to be detected and size and the relative position of each parts in position and the window area, neighborhood characteristics is represented the image difference of the neighborhood of window area and window area; Work in coordination with and the character representation window area occurs and have the relation of the window area of maximum confidence.
Alternatively, the contextual feature of window area is represented by vector f:
f＝(σ(sc)，r，p，q，σ(s m)，r m)
Wherein, σ (sc)=1/ (1+exp (2sc)),
Wherein, sc is the degree of confidence of window area, and r represents position and the size of window area, and p represents that each parts in the window area are with respect to the position at root area center, q represents the gradation of image mean difference of the adjacent area of specific region in the window area and window area, s mBe the maximum confidence in the degree of confidence of described a plurality of window areas, r mBe position and the size with window area of maximum confidence.
Alternatively, described object detection systems also comprises: context deformable component model detecting unit, to from described a plurality of window areas, remove the pseudo-window area contextual feature input context sorter of remaining window area afterwards, to obtain the new degree of confidence of window area, wherein, the context sorter is to utilize as the contextual feature of sample to train the sorter that obtains.
According to a further aspect in the invention, provide a kind of object detection method, comprising: receive image to be detected; Utilize the root template to treat detected image and scan, to extract the characteristics of image of a plurality of window areas;
By the characteristics of image input deformable component model of a plurality of window areas that will extract, obtain the degree of confidence of described a plurality of window areas to utilize the deformable component model, wherein, the size of adjusting each window area makes the degree of confidence maximum of this window area; Determine to exist the window area of object according to the degree of confidence of window area.
Alternatively, adjust the scope of each window area that extracts, the corresponding convolution filter of each window area after feasible the adjustment reaches maximum with the dot product of the characteristics of image on each window area after the adjustment, makes the degree of confidence of each window area reach maximum.
Alternatively, the deformable component model obtains by training, wherein, when training deformable component model, adjusts the size as the window area of sample, makes that the degree of confidence as the window area of sample reaches maximum.
Alternatively, when training deformable component model, adjustment is as the scope of the window area of sample, the corresponding convolution filter of the window area as sample after make adjusting with reach maximum at the dot product as the characteristics of image on the window area of sample after the adjustment, make that the degree of confidence as the window area of sample reaches maximum.
Alternatively, described deformable component model is to mix the deformable component model.
Alternatively, described method also comprises: remove pseudo-window area according to the interactive relation between described a plurality of window areas from the described a plurality of window areas that obtained degree of confidence.
Alternatively, the step of removing pseudo-window area comprises: the characteristic information extraction unit, from each window area characteristic information extraction; The redundant unit of removing utilizes the characteristic information that extracts to determine described interactive relation, to remove pseudo-window area from described a plurality of window areas.
Alternatively, described characteristic information comprises at least one in the degree of confidence of degree of confidence, parts of the positional information of the positional information of degree of confidence, root of window area and/or yardstick information, parts and/or yardstick information, root.
Alternatively, judge and remove pseudo-window area by maximizing following equation:
Figure BSA00000648239000051
Wherein, M represents the quantity of described a plurality of window areas;
φ (x i, y i)=y iX i
Figure BSA00000648239000052
x i=(v i(s), Z), v i(s) degree of confidence of i window area of expression, Z represents the vector of K dimension, K represents the quantity of the deformable component model that described mixing deformable component model is included, the v of Z i(c) individual element is that other elements of 1, Z are zero, v i(c) expression detects the index of i the employed deformable component model of window area; y iThat represents i window area is used for whether sign is the binary score of pseudo-window area; y jWhether what expression was used for j window area is the binary score of pseudo-window area for sign;
Figure BSA00000648239000053
The representation model parameter, d IjRepresent the interactive relation between i window area and j the window area, wherein, when described equation maximization, the window area with binary score of the pseudo-window area of sign is judged as pseudo-window area.
Alternatively, use the φ (x of precognition i, y i),
Figure BSA00000648239000061
SS divides class methods by predetermined structure to be trained to obtain
Figure BSA00000648239000062
Alternatively, the interactive relation between the window area comprise that root-root is mutual, root-parts mutual, at least one in mutual of parts-parts.
Alternatively, described method also comprises: the contextual feature input context sorter that will obtain the window area of degree of confidence, to obtain the new degree of confidence of window area, wherein, the context sorter is to utilize as the contextual feature of sample to train the sorter that obtains.
Alternatively, contextual feature comprises: shaped position feature, neighborhood characteristics, the collaborative feature that occurs.
Alternatively, the size of shaped position character representation window area in image to be detected and size and the relative position of each parts in position and the window area, neighborhood characteristics is represented the image difference of the neighborhood of window area and window area; Work in coordination with and the character representation window area occurs and have the relation of the window area of maximum confidence.
Alternatively, the contextual feature of window area is represented by vector f:
f＝(σ(sc)，r，p，q，σ(s m)，r m)，
Wherein, σ (sc)=1/ (1+exp (2sc)),
Wherein, sc is the degree of confidence of window area, and r represents position and the size of window area, and p represents that each parts in the window area are with respect to the position at root area center, q represents the gradation of image mean difference of the adjacent area of specific region in the window area and window area, s mBe the maximum confidence in the degree of confidence of described a plurality of window areas, r mBe position and the size with window area of maximum confidence.
Alternatively, described method also comprises: will remove the pseudo-window area contextual feature input context sorter of remaining window area afterwards from described a plurality of window areas, to obtain the new degree of confidence of window area, wherein, the context sorter is to utilize as the contextual feature of sample to train the sorter that obtains.
Another aspect of the present invention provides a kind of object detection systems, comprising: image receiving unit receives image to be detected; Feature extraction unit is extracted the characteristics of image of a plurality of window areas; Deformable component model detecting unit, the characteristics of image input deformable component model by a plurality of window areas that will extract obtains the degree of confidence of described a plurality of window areas to utilize the deformable component model; The object determining unit is determined to exist the window area of object according to the degree of confidence of window area.
Alternatively, described object detection systems also comprises: the redundant unit that suppresses, from the described a plurality of window areas that obtained degree of confidence, remove pseudo-window area according to the interactive relation between described a plurality of window areas, wherein, the redundant unit that suppresses comprises: the characteristic information extraction unit, from each window area characteristic information extraction; The redundant unit of removing utilizes the characteristic information that extracts to determine described interactive relation, to remove pseudo-window area from described a plurality of window areas.
Alternatively, described object detection systems also comprises: context deformable component model detecting unit, to from described a plurality of window areas, remove the pseudo-window area contextual feature input context sorter of remaining window area afterwards, to obtain the new degree of confidence of window area, wherein, the context sorter is to utilize as the contextual feature of sample to train the sorter that obtains.
Another aspect of the present invention provides a kind of object detection method, comprising: receive image to be detected; Extract the characteristics of image of a plurality of window areas; Characteristics of image input deformable component model by a plurality of window areas that will extract obtains the degree of confidence of described a plurality of window areas to utilize the deformable component model; Determine to exist the window area of object according to the degree of confidence of window area.
Alternatively, described object detection method also comprises: remove pseudo-window area according to the interactive relation between described a plurality of window areas from the described a plurality of window areas that obtained degree of confidence, wherein, removing pseudo-window area comprises: from each window area characteristic information extraction; Utilize the characteristic information that extracts to determine described interactive relation, from described a plurality of window areas, to remove pseudo-window area.
Alternatively, described object detection method also comprises: will remove the pseudo-window area contextual feature input context sorter of remaining window area afterwards from described a plurality of window areas, to obtain the new degree of confidence of window area, wherein, the context sorter is to utilize as the contextual feature of sample to train the sorter that obtains.
According to technical scheme of the present invention, by improving the data item in root definition among the existing DPM, can effectively overcome and have the object of difference among a small circle with the template length breadth ratio and be easy to the problem of in detection, being lost.
In addition, according to Redundancy-Restraining Technique of the present invention, solved because object is blocked, combine with other objects and the score of the window area that object space layout and overlapping complicacy cause is detected by DPM may inaccurate problem, effectively eliminated pseudo-window area.
In addition, according to the present invention, utilize contextual feature that the classification score of window area is proofreaied and correct, can further improve accuracy of detection, particularly improved the accuracy of detection to the object in the medical image.
Will be in ensuing description part set forth the present invention other aspect and/or advantage, some will be clearly by describing, and perhaps can learn through enforcement of the present invention.
Description of drawings
By the detailed description of carrying out below in conjunction with accompanying drawing, above and other objects of the present invention, characteristics and advantage will become apparent, wherein:
Fig. 1 is the block diagram that illustrates according to the object detection systems of the object in the detected image of the embodiment of the invention;
Fig. 2 illustrates the block diagram that suppresses the unit according to the redundancy of the embodiment of the invention.
Fig. 3 is the process flow diagram that the object detection method of the object in the detected image according to an embodiment of the invention is shown.
Fig. 4 is the process flow diagram of the object detection method of the object in the detected image that illustrates according to another embodiment of the present invention.
Embodiment
Below, exemplary embodiment of the present invention is described with reference to the accompanying drawings more fully, exemplary embodiment is shown in the drawings.Run through the description to accompanying drawing, identical label is represented components identical.
Fig. 1 is the block diagram that illustrates according to the object detection systems 100 of the object in the detected image of the embodiment of the invention.
Detection system 100 comprises image receiving unit 110, feature extraction unit 120, DPM detecting unit 130, object determining unit 140.
Image receiving unit 110 is used for receiving image to be detected.
Feature extraction unit 120 is used for extracting characteristics of image from image to be detected.
The characteristics of image that extracts can be that for example gradient orientation histogram (HOG) feature, local binary (LBP) feature, trellis depth feature (GDF), yardstick invariant features are changed various characteristics of image such as (SIFT) feature.
Feature extraction unit 120 can be used the root template to treat detected image and scan, thereby obtains a plurality of window areas (that is a plurality of roots) and characteristics of image thereof.
DPM detecting unit 130 can use the DPM of training in advance to come detected object.DPM detecting unit 130 is imported DPM with the characteristics of image of each window area that feature extraction unit 120 is extracted, thereby obtains the classification score (that is degree of confidence) of each window area.
DPM according to the present invention improves existing DPM.
In existing DPM, according to equation (1) as can be known, at the data item of root definition (that is convolution filter F corresponding with root (the also window area that namely extracts), 0With at root P 0The dot product of the feature of extracting) can be expressed as followsin: F 0. φ (h, p 0).
In DPM according to the present invention, the data item that defines at root is expressed as followsin:
max τ ( F 0 ( τ ) · φ ( H , p 0 , τ ) ) - - - ( 4 )
Wherein,
Figure BSA00000648239000092
P 0The regional extent of expression root, F 0(τ) the convolution filter F of expression root correspondence 0Part on regional τ, φ (H, p 0, τ) expression by characteristics of image pyramid H obtain at P 0The part of root feature on regional τ.
Equation (4) expression is adjusted into the feasible convolution filter F that is applied on the regional τ of τ by the scope with original root area 0(τ) and the dot product of the proper vector of extracting at regional τ (that is the data item of the definition of the root after adjustment) maximum.Preferably, consider computational complexity and regional representativeness, the variation range of variable τ is restricted to several and is equal to or slightly less than regional P 0The rectangular area.
At this moment, equation 1 can be rewritten as:
f ( p 0 , . . . p n ) = max τ ( F 0 ( τ ) · φ ( H , p 0 , τ ) ) + Σ i = 1 n F i · φ ( H , p i ) - Σ i = 1 n d i · φ d ( dx i , dy i ) + b - - - ( 5 )
Specifically, when training DPM, adjust the root P as sample 0The scope of region make with adjust after root P 0Corresponding convolution filter F 0The dot product maximum of the feature of extracting with root after adjustment.When using DPM to detect, adjust the root P as input 0The scope of region make with adjust after root P 0Corresponding convolution filter F 0The dot product maximum of the feature of extracting with root after adjustment.According to equation (5) as can be known, with root P 0Corresponding convolution filter F 0With the dot product maximum of the feature of extracting at root, it also is final classification score maximum.
Object determining unit 140 determines to exist the window of object according to the classification score (that is degree of confidence) of window area.Should be appreciated that this neighborhood technician is under the situation of the classification score that obtains window area, it is known determining to exist the technology of the window area of object.For example, can be with the window area of the classification score maximum zone as the object place; Perhaps when the classification score of window area during less than predetermined threshold, determine in this window area, not exist object; When the classification score of window area during more than or equal to predetermined threshold, determine in this window area, to have object.
As implied above, improve by the data item in root definition, can eliminate effectively and have the object of difference among a small circle with the template length breadth ratio and be easy to the problem of in detection, being lost.
In a further embodiment, DPM detecting unit 130 uses and mixes DPM.Mix DPM and be made of a plurality of DPM, the root template of each DPM differs from one another.At this moment, feature extraction unit 120 is used the root template of each DPM to treat detected image respectively and is scanned, thereby obtains a plurality of window areas and characteristics of image thereof.
The characteristics of image of each in described a plurality of window area is transfused to mixing DPM, thereby obtains the classification score of each window area.
In some specific occasions, owing to object is blocked, combines with other objects and object space layout and overlapping complicacy, therefore the score of the window area of DPM detection may be inaccurate, some pseudo-window areas (namely, the window area that does not have object) possible score is higher, and this has greatly reduced accuracy of detection.In order to address this problem, the present invention proposes a kind of Redundancy-Restraining Technique, get rid of these pseudo-window areas, thus can according to from described a plurality of window areas, got rid of classifying of window area after the pseudo-window area assign to determine to exist the window area of object.
According to another embodiment of the present invention, object detection systems 100 also comprises the redundant unit (not shown) that suppresses.Pseudo-window area can be removed in the redundant unit that suppresses from the window area of DPM detecting unit 130 outputs.In other words, redundantly suppress pseudo-window area can be removed in the unit from the classification score of a plurality of window areas of DPM detecting unit 130 outputs classification score.
Fig. 2 illustrates the block diagram that suppresses the unit according to the redundancy of the embodiment of the invention.
The redundant unit that suppresses comprises characteristic information extraction unit 141, the redundant unit 142 of removing.
Characteristic information extraction unit 141 is from each window area corresponding window area of classification score of output (that is, with) characteristic information extraction.Specifically, described characteristic information can comprise at least one in the score information of yardstick information, root and parts of positional information, root and parts of the root of skew that the model of PTS, the use of distortion cost, the window area of the parts of object causes, the object in the window area and parts.
For example, the characteristic information from i window area can be represented as v i,
v i＝(b，s，s 0，s 1，...s D，dd 1，..dd D，l 0，l 1...l D，c)， (6)
Wherein, l 0Be position and the yardstick of the root of object, l 1... l DBe position and the yardstick of the parts of object, the quantity of the parts of D indicated object, s 0Be the score of root, s 1... s DBe the score of parts, dd 1..dd DBe the distortion cost of parts, the skew that the DPM among the mixing DPM that b is to use causes, s is the PTS of window area, c be object component index (namely, represent that i window area obtained by c the DPM detection that mixes among the DPM), 1≤c≤K, K represent to mix the quantity of the DPM among the DPM.
Should be appreciated that, although top v iIn comprised multiple information, but should be appreciated that, can only extract partial information wherein as required.
The redundant unit 142 of removing utilizes the characteristic information that extracts to determine from the interactive relation between the window area of DPM detecting unit 130 outputs, to remove pseudo-window area from the window area of DPM detecting unit 130 outputs.This interactive relation has embodied the overlapping characteristic between the window area.
Specifically, suppose x iBe the characteristic information that extracts from window area i, then entire image can be represented as the characteristic information X={x of extraction i: i=1...M}, M represent the quantity of window area.If being carried out binary, each window area marks to determine whether it is correct example, the then mark of i window area: y i{ 0,1} (should be appreciated that bi-values of the present invention is not limited to 0 and 1, also can use other value as bi-values) makes Y={y to ∈ i: i=1...M}, use must being divided into of vectorial Y mark image X:
Figure BSA00000648239000111
Wherein, φ (x i, y i)=y iX i, The representation model parameter, x i=(v i(s), Z), Z represents the vector of K dimension, the v of Z i(c) individual element is 1, and other elements are zero, d IjRepresent the interactive relation (that is overlapping relation) between i window area and the j window area.
Can use the φ (x of precognition i, y i), SS trains to obtain by existing structured sorting technique (for example, structuring SVM algorithm, Boost algorithm etc.)
Figure BSA00000648239000115
Preferably, use structuring SVM algorithm.Owing to use the structuring sorting technique to obtain
Figure BSA00000648239000116
Be existing technology, will no longer describe in detail.
According to embodiments of the invention, the interactive relation between the different windows zone comprises that root-root is mutual, root-parts mutual, at least one in mutual of parts-parts.
Root-root has embodied the overlapping characteristic (for example, the overlapping characteristic between the root of the root of a window area and another window area) between the root in different windows zone alternately.
Root-parts have embodied root and the overlapping characteristic between the parts (for example, the overlapping characteristic between the parts of the root of a window area and another window area) in different windows zone alternately.
Parts-parts have embodied the overlapping characteristic (for example, the overlapping characteristic between the parts of the parts of a window area and another window area) between the parts in different windows zone alternately.
Root-root is mutual, root-parts are mutual, parts-parts can be represented as respectively alternately
Figure BSA00000648239000117
Know
Figure BSA00000648239000118
For example, when using all above-mentioned three kinds when mutual,
Figure BSA00000648239000119
Figure BSA000006482390001110
Represent the interactive relation between the root between i window area and the j window area.Be appreciated that
Figure BSA000006482390001111
It is the matrix of K * K.In one example, the arbitrary element in this matrix
Figure BSA000006482390001112
(m, n represent the index of the element in this matrix, for example, the capable n row of m) can be expressed as following equation (8):
Figure BSA000006482390001113
Here, ol (v i(l 0), v i(l 0)) Duplication between the root of expression i window area and the root of j window area.
Represent root and the interactive relation between the parts (that is the interactive relation between the parts of the root of i window area and j window area) between i window area and the j window area.Be appreciated that
Figure BSA00000648239000122
Be K * (matrix of K * D), arbitrary element in this matrix
Figure BSA00000648239000123
Can be expressed as following equation (9):
Here, g ∈ [1, D], ol (v i(l 0), v i(l g)) Duplication between the root of expression i window area and g the parts of j window area.
Represent parts between i window area and the j window area and the interactive relation between the parts.Be appreciated that
Figure BSA00000648239000126
Be (K * D) * (the matrix of K * D).In one example, the arbitrary element in this matrix
Figure BSA00000648239000127
Can be expressed as following equation (10):
....(10)
Here, e ∈ [1, D], g ∈ [1, D], ol (v i(l e), v j(l g)) Duplication between e parts of expression i window area and g the parts of j window area.
Calculating makes equation 7 be maximum Y, and this calculating can be represented as argmax YS (X, Y).At this moment, be noted as 1 window area and be considered to the final example that detects, be noted as 0 window area and be considered to pseudo-window area.
Can make to calculate in various manners to make equation 7 for maximum Y, for example, can use the mode of enumerating.
In another embodiment of the present invention, using greedy algorithm to calculate makes equation 7 be maximum Y.
In according to another embodiment of the present invention, detection system 100 also comprises context DPM detecting unit (not shown).Context DPM detecting unit is proofreaied and correct DPM detecting unit 130 or the redundant classification score that suppresses each window area of unit output according to the contextual feature of window area.
Specifically, the contextual feature input context sorter of the window area that context DPM detecting unit will be corresponding with the classification score of DPM detecting unit 130 or redundant inhibition unit output obtains the new score of window area.At this moment, object determining unit 140 is classified to such an extent that assign to determine to exist the window of object according to window area new.
The context sorter is to utilize as the contextual feature of sample to train the sorter that obtains.Preferably, from being used for training DPM or mixing the sample extraction contextual feature of DPM as the contextual feature of sample, can improve training speed and precision like this.
Contextual feature comprises: the shaped position feature; Neighborhood characteristics; The collaborative feature that occurs.The size of shaped position character representation window area in image to be detected and size and the relative position of each parts in position and the window area.Neighborhood characteristics is represented the image difference of the neighborhood of window area and window area.Preferably, the area of window area is identical with the area of the neighborhood of window area.For example, image difference can be the average image gray scale difference or gray variance and the position-statistic such as gray scale covariance of root area and component area and described neighborhood.The collaborative relation that the window area of maximum score occurs having in character representation window area and detected all window areas.For example, if current window area is not the window area with maximum score, then collaboratively feature occurs and in the context sorter, generally can inhibiting effect be arranged to current window area.
In one example, the contextual feature of a window area can be represented by following vector f:
f＝(σ(sc)，r，p，q，σ(s m)，r m)
Wherein, σ (sc)=1/ (1+exp (2sc)),
Wherein, sc is the score of window area, and r represents position and the size of window area, and p represents that each parts in the window area are with respect to the position at root area center, q represents the gradation of image mean difference of the adjacent area of specific region in the window area and window area, s mMaximum score in the score of window area, r mBe position and the size with window area of described maximum score.
During the neoplastic lesion in detecting medical image etc., compare with other object detection, it is big that tumour in the medical image has change of shape, poor contrast, characteristics such as noise is obvious, more particularly the characteristics of image of a part of tumor region has ambiguity, that is to say, an almost identical image block, in piece image, can be considered to tumour, in another width of cloth image, will be considered to not be tumour, or even can be considered to tumour in a position appearance of same image, and tumour occur to be considered to not be in another position.By contextual feature of the present invention the classification score is proofreaied and correct, can improve the accuracy of detection of neoplastic lesion object effectively.
In addition, in a further embodiment, the redundant unit, context DPM detecting unit of suppressing according to the present invention can be separately or be applied to together (for example, utilize the object detection systems of the DPM of equation 1) in the existing object detection systems based on DPM.
Fig. 3 illustrates the process flow diagram of the object detection method of the object in the detected image according to an embodiment of the invention.
In step 301, receive image to be detected.
In step 302, extract characteristics of image from image to be detected.Specifically, utilize the root template to treat detected image and scan, thereby obtain a plurality of window areas and characteristics of image thereof.
In step 303, will import the DPM of training in advance at the characteristics of image that step 302 is extracted, thereby obtain the classification score of each window area.
In step 304, according to classifying of the window area that obtains in step 303 assign to determine to exist the window area of object.
Fig. 4 illustrates the process flow diagram of the object detection method of the object in according to another embodiment of the present invention the detected image.
In step 401, receive image to be detected.
In step 402, extract characteristics of image from image to be detected.Specifically, utilize the root template to treat detected image and scan, thereby obtain a plurality of window areas and characteristics of image thereof.
In step 403, will import the DPM of training in advance at the characteristics of image that step 402 is extracted, thereby obtain the classification score of each window area.
In step 404, according to classifying of the window area that obtains in step 403 assign to determine to exist the window area of object.
In step 405, from each window area characteristic information extraction.Described characteristic information can comprise at least one in the score of score, parts of the positional information of the positional information of PTS, root of window area and/or yardstick information, parts and/or yardstick information, root.
In step 406, utilize the characteristic information that extracts from described a plurality of window areas, to remove pseudo-window area.Specifically, the result of maximization formula (7), thus the window area with binary score of the pseudo-window area of sign is judged as pseudo-window area.
In a further embodiment, after step 303,403 or 406, also can comprise step: extract the contextual feature of window area, and contextual feature is imported the new score that the context sorter obtains window area.Should be appreciated that the window area that is extracted contextual feature is not included in the pseudo-window area that step 405 is determined.
According to the present invention, by improving the account form in the data item of root definition among the existing DPM, can effectively overcome and have the object of difference among a small circle with the template length breadth ratio and be easy to the problem of in detection, being lost.In addition, according to Redundancy-Restraining Technique of the present invention, solved because object is blocked, combine with other objects and the score of the window area that object space layout and overlapping complicacy cause is detected by DPM may inaccurate problem, effectively eliminated pseudo-window area.In addition, according to the present invention, utilize contextual feature that the classification score of window area is proofreaied and correct, can further improve accuracy of detection, particularly improved the accuracy of detection to the object in the medical image.
Although specifically shown with reference to its exemplary embodiment and described the present invention, but the technician of this neighborhood should be appreciated that, under the situation that does not break away from the spirit and scope of the present invention that claim limits, can carry out various changes on form and the details to it.

Claims (16)

1.一种对象检测系统，包括：1. An object detection system comprising: 图像接收单元，接收待检测图像；The image receiving unit receives the image to be detected; 特征提取单元，利用根部模板对待检测图像进行扫描，以提取多个窗口区域的图像特征；The feature extraction unit scans the image to be detected by using the root template to extract image features of multiple window regions; 可变形部件模型检测单元，通过将提取的多个窗口区域的图像特征输入可变形部件模型，以利用可变形部件模型获得所述多个窗口区域的置信度，其中，可变形部件模型调整每个窗口区域的大小使得每个窗口区域的置信度达到最大；The deformable part model detection unit is configured to use the deformable part model to obtain the confidence of the multiple window regions by inputting the extracted image features of the multiple window regions into the deformable part model, wherein the deformable part model adjusts each The size of the window area is such that the confidence of each window area is maximized; 对象确定单元，根据窗口区域的置信度确定存在对象的窗口区域。The object determination unit determines the window area where the object exists according to the confidence of the window area. 2.根据权利要求1所述的对象检测系统，其中，调整提取的每个窗口区域的范围，使得调整后的每个窗口区域所对应的卷积滤波器与在调整后的每个窗口区域上的图像特征的点积达到最大，来使得每个窗口区域的置信度达到最大。2. The object detection system according to claim 1, wherein the range of each window region extracted is adjusted such that the convolution filter corresponding to each window region after adjustment is the same as that on each window region after adjustment The dot product of the image features is maximized to maximize the confidence of each window region. 3.根据权利要求1所述的对象检测系统，其中，可变形部件模型是通过训练得到的，其中，在训练可变形部件模型时，调整作为样本的窗口区域的大小，使得作为样本的窗口区域的置信度达到最大。3. The object detection system according to claim 1, wherein the deformable part model is obtained through training, wherein, when training the deformable part model, the size of the window area as a sample is adjusted so that the window area as a sample confidence reaches the maximum. 4.根据权利要求3所述的对象检测系统，其中，在训练可变形部件模型时，调整作为样本的窗口区域的范围，使得调整后的作为样本的窗口区域所对应的卷积滤波器与在调整后的作为样本的窗口区域上的图像特征的点积达到最大，来使得作为样本的窗口区域的置信度达到最大。4. The object detection system according to claim 3, wherein, when training the deformable part model, the range of the window area as a sample is adjusted so that the adjusted convolution filter corresponding to the window area as a sample is the same as that in The adjusted dot product of the image feature on the window area serving as the sample reaches the maximum, so that the confidence of the window area serving as the sample reaches the maximum. 5.根据权利要求1所述的对象检测系统，其中，所述可变形部件模型是混合可变形部件模型。5. The object detection system of claim 1, wherein the deformable part model is a hybrid deformable part model. 6.根据权利要求5所述的对象检测系统，还包括：冗余抑制单元，根据所述多个窗口区域之间的交互关系从获得了置信度的所述多个窗口区域中去除伪窗口区域，6. The object detection system according to claim 5, further comprising: a redundancy suppression unit, which removes false window regions from the plurality of window regions for which the confidence is obtained according to the interaction relationship between the plurality of window regions , 其中，冗余抑制单元包括：Among them, the redundant suppression unit includes: 特征信息提取单元，从每个窗口区域提取特征信息；A feature information extraction unit extracts feature information from each window area; 冗余去除单元，利用提取的特征信息确定所述交互关系，以从所述多个窗口区域中去除伪窗口区域。The redundancy removing unit determines the interaction relationship by using the extracted feature information, so as to remove the pseudo window area from the plurality of window areas. 7.根据权利要求6所述的对象检测系统，其中，所述特征信息包括窗口区域的置信度、根部的位置信息和/或尺度信息、部件的位置信息和/或尺度信息、根部的置信度、部件的置信度中的至少一个。7. The object detection system according to claim 6, wherein the feature information comprises confidence of the window area, position information and/or scale information of the root, position information and/or scale information of the part, confidence of the root , at least one of the confidence of the component. 8.根据权利要求6所述的对象检测系统，其中，冗余去除单元通过最大化下面的等式来判定并去除伪窗口区域：8. The object detection system according to claim 6, wherein the redundancy removal unit determines and removes the false window region by maximizing the following equation:
Figure FSA00000648238900021
Figure FSA00000648238900021
 其中，M表示所述多个窗口区域的数量；Wherein, M represents the number of the plurality of window regions; φ(xi，yi)＝yi·xi；
Figure FSA00000648238900022
xi＝(vi(s)，Z)，vi(s)表示第i个窗口区域的置信度，Z表示K维的向量，K表示所述混合可变形部件模型所包括的可变形部件模型的数量，Z的第vi(c)个元素为1，Z的其他元素为零，vi(c)表示检测出第i个窗口区域所使用的可变形部件模型的索引；yi表示第i个窗口区域的用于标识是否是伪窗口区域的二元得分；yi表示用于第j个窗口区域的用于标识是否是伪窗口区域的二元得分；
Figure FSA00000648238900023
表示模型参数，dij表示第i个窗口区域和第j个窗口区域之间的交互关系，φ(x i , y i )=y i x i ;
Figure FSA00000648238900022
x i = (v i (s), Z), v i (s) represents the confidence of the i-th window area, Z represents a K-dimensional vector, and K represents the deformable components included in the hybrid deformable component model The number of models, the v i (c) element of Z is 1, and the other elements of Z are zero, v i (c) represents the index of the deformable component model used to detect the i-th window area; y i represents The i-th window area is used to identify whether it is a binary score of a pseudo-window area; y i represents the binary score used to identify whether it is a pseudo-window area for the j-th window area;
Figure FSA00000648238900023
Indicates the model parameters, d ij indicates the interaction relationship between the i-th window area and the j-th window area, 其中，当所述等式最大化时，具有标识伪窗口区域的二元得分的窗口区域被判定为伪窗口区域。Wherein, when said equation is maximized, a window region having a binary score identifying a pseudo window region is determined to be a pseudo window region. 9.根据权利要求8所述的对象检测系统，其中，使用预知的φ(xi，yi)、
Figure FSA00000648238900024
SS通过预定结构化分类方法进行训练来获得
Figure FSA00000648238900025
9. The object detection system according to claim 8, wherein using the pre-knowledged φ(x i , y i ),
Figure FSA00000648238900024
SS is obtained by training with a predetermined structured classification method
Figure FSA00000648238900025
 10.根据权利要求6所述的对象检测系统，其中，窗口区域之间的交互关系包括根部-根部交互、根部-部件交互、部件-部件交互中的至少一个。10. The object detection system according to claim 6, wherein the interaction relationship between window regions comprises at least one of root-root interaction, root-component interaction, component-component interaction. 11.根据权利要求1所述的对象检测系统，还包括：上下文可变形部件模型检测单元，将获得了置信度的窗口区域的上下文特征输入上下文分类器，以获得窗口区域的新的置信度，其中，上下文分类器是利用作为样本的上下文特征训练得到的分类器。11. The object detection system according to claim 1, further comprising: a context deformable part model detection unit, which inputs the context feature of the window region with confidence into the context classifier to obtain a new confidence of the window region, Wherein, the context classifier is a classifier trained by using context features as samples. 12.根据权利要求11所述的对象检测系统，其中，上下文特征包括：形状位置特征、邻域特征、协同出现特征。12. The object detection system according to claim 11, wherein the context features include: shape position features, neighborhood features, co-occurrence features. 13.根据权利要求11所述的对象检测系统，其中，形状位置特征表示窗口区域在待检测图像中的大小和位置以及窗口区域中的各部件的大小和相对位置，邻域特征表示窗口区域与窗口区域的邻域的图像差异；协同出现特征表示窗口区域与具有最大置信度的窗口区域的关系。13. The object detection system according to claim 11, wherein the shape position feature represents the size and position of the window area in the image to be detected and the size and relative position of each component in the window area, and the neighborhood feature represents the relationship between the window area and Image difference of the neighborhood of the window region; co-occurrence features represent the relationship of the window region to the window region with the greatest confidence. 14.根据权利要求12所述的对象检测系统，其中，窗口区域的上下文特征由向量f表示：14. The object detection system according to claim 12, wherein the contextual features of the window region are represented by the vector f: f＝(σ(sc)，r，p，q，σ(sm)，rm)f = (σ(sc), r, p, q, σ(s m ), r m ) 其中，σ(sc)＝1/(1+exp(-2sc))，where, σ(sc)=1/(1+exp(-2sc)), 其中，sc是窗口区域的置信度，r表示窗口区域的位置和大小，p表示窗口区域中的每个部件相对于根部区域中心的位置，q表示窗口区域内的特定区域与窗口区域的相邻区域的图像灰度平均差，sm是所述多个窗口区域的置信度中的最大置信度，rm是具有最大置信度的窗口区域的位置和大小。Among them, sc is the confidence degree of the window area, r indicates the position and size of the window area, p indicates the position of each component in the window area relative to the center of the root area, and q indicates the adjacency of a specific area in the window area to the window area The image gray level average difference of the region, s m is the maximum confidence degree among the confidence degrees of the plurality of window regions, and rm is the position and size of the window region with the maximum confidence degree. 15.根据权利要求4所述的对象检测系统，还包括：上下文可变形部件模型检测单元，将从所述多个窗口区域中去除伪窗口区域之后剩余的窗口区域的上下文特征输入上下文分类器，以获得窗口区域的新的置信度，其中，上下文分类器是利用作为样本的上下文特征训练得到的分类器。15. The object detection system according to claim 4, further comprising: a context deformable part model detection unit, which inputs the context features of the remaining window regions after removing the pseudo window regions from the plurality of window regions to the context classifier, To obtain a new confidence level of the window region, wherein the context classifier is a classifier trained by using the context feature as a sample. 16.一种对象检测方法，包括：16. A method of object detection comprising: 接收待检测图像；Receive the image to be detected; 利用根部模板对待检测图像进行扫描，以提取多个窗口区域的图像特征；Use the root template to scan the image to be detected to extract image features of multiple window regions; 通过将提取的多个窗口区域的图像特征输入可变形部件模型，以利用可变形部件模型获得所述多个窗口区域的置信度，其中，调整每个窗口区域的大小使得该窗口区域的置信度最大；By inputting the extracted image features of a plurality of window regions into the deformable part model, the confidence of the plurality of window regions is obtained by using the deformable part model, wherein the size of each window region is adjusted so that the confidence of the window region maximum; 根据窗口区域的置信度确定存在对象的窗口区域。The window area where the object exists is determined according to the confidence of the window area.
CN2011104567794A 2011-12-27 2011-12-27 Object detecting system and object detecting method Pending CN103186790A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN2011104567794A CN103186790A (en) 2011-12-27 2011-12-27 Object detecting system and object detecting method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN2011104567794A CN103186790A (en) 2011-12-27 2011-12-27 Object detecting system and object detecting method

Publications (1)

Publication Number Publication Date
CN103186790A true CN103186790A (en) 2013-07-03

Family

ID=48677950

Family Applications (1)

Application Number Title Priority Date Filing Date
CN2011104567794A Pending CN103186790A (en) 2011-12-27 2011-12-27 Object detecting system and object detecting method

Country Status (1)

Country Link
CN (1) CN103186790A (en)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103971134A (en) * 2014-04-25 2014-08-06 华为技术有限公司 Image classifying, retrieving and correcting method and corresponding device
CN104200236A (en) * 2014-08-22 2014-12-10 浙江生辉照明有限公司 Quick target detection method based on DPM (deformable part model)
CN104484680A (en) * 2014-09-26 2015-04-01 徐晓晖 Multi-model multi-threshold combined pedestrian detection method
CN105900084A (en) * 2013-12-20 2016-08-24 高通股份有限公司 System, method and device for image retrieval
CN106326891A (en) * 2015-06-30 2017-01-11 展讯通信(天津)有限公司 Mobile terminal, target detection method and device of mobile terminal
CN106778665A (en) * 2016-12-29 2017-05-31 浙江大华技术股份有限公司 A kind of vehicle window detection method and device
CN107851192A (en) * 2015-05-13 2018-03-27 北京市商汤科技开发有限公司 For detecting the apparatus and method of face part and face
CN108229495A (en) * 2017-06-23 2018-06-29 北京市商汤科技开发有限公司 Target object detection method, device, electronic equipment and storage medium
CN108388874A (en) * 2018-03-05 2018-08-10 厦门大学 Prawn morphological parameters method for automatic measurement based on image recognition and cascade classifier
CN108830210A (en) * 2018-06-11 2018-11-16 广东美的制冷设备有限公司 Human body detecting method and device based on image
CN109977965A (en) * 2019-02-28 2019-07-05 北方工业大学 A method and device for determining detection targets in remote sensing airport images

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101872477A (en) * 2009-04-24 2010-10-27 索尼株式会社 Method and device for detecting object in image and system containing device
CN102129569A (en) * 2010-01-20 2011-07-20 三星电子株式会社 Equipment and method for detecting object based on multiscale comparison characteristic

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101872477A (en) * 2009-04-24 2010-10-27 索尼株式会社 Method and device for detecting object in image and system containing device
CN102129569A (en) * 2010-01-20 2011-07-20 三星电子株式会社 Equipment and method for detecting object based on multiscale comparison characteristic

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
PEDRO F. FELZENSZWALB ET AL.: ""Object Detection with Discriminatively Trained Part-Based Models"", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 *
安国成 等: ""多窗口目标跟踪算法"", 《计算机研究与发展》 *
胡正平,杨建秀: ""HOG特征混合模型结合隐SVM的感兴趣目标检测定位算法"", 《信号处理》 *

Cited By (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105900084A (en) * 2013-12-20 2016-08-24 高通股份有限公司 System, method and device for image retrieval
CN105900084B (en) * 2013-12-20 2019-12-31 高通股份有限公司 System, method and device for image retrieval
US10346465B2 (en) 2013-12-20 2019-07-09 Qualcomm Incorporated Systems, methods, and apparatus for digital composition and/or retrieval
CN103971134A (en) * 2014-04-25 2014-08-06 华为技术有限公司 Image classifying, retrieving and correcting method and corresponding device
CN103971134B (en) * 2014-04-25 2017-07-07 华为技术有限公司 Image classification, retrieval and bearing calibration, and related device
CN104200236B (en) * 2014-08-22 2018-10-26 浙江生辉照明有限公司 Fast target detection method based on DPM
CN104200236A (en) * 2014-08-22 2014-12-10 浙江生辉照明有限公司 Quick target detection method based on DPM (deformable part model)
WO2016026371A1 (en) * 2014-08-22 2016-02-25 Zhejiang Shenghui Lighting Co., Ltd. Fast object detection method based on deformable part model (dpm)
EP3183691A4 (en) * 2014-08-22 2017-11-08 Zhejiang Shenghui Lighting Co., Ltd Fast object detection method based on deformable part model (dpm)
US9846821B2 (en) 2014-08-22 2017-12-19 Zhejiang Shenghui Lighting Co., Ltd Fast object detection method based on deformable part model (DPM)
CN104484680A (en) * 2014-09-26 2015-04-01 徐晓晖 Multi-model multi-threshold combined pedestrian detection method
CN107851192A (en) * 2015-05-13 2018-03-27 北京市商汤科技开发有限公司 For detecting the apparatus and method of face part and face
CN106326891A (en) * 2015-06-30 2017-01-11 展讯通信(天津)有限公司 Mobile terminal, target detection method and device of mobile terminal
CN106778665B (en) * 2016-12-29 2019-09-17 浙江大华技术股份有限公司 A kind of vehicle window detection method and device
CN106778665A (en) * 2016-12-29 2017-05-31 浙江大华技术股份有限公司 A kind of vehicle window detection method and device
CN108229495A (en) * 2017-06-23 2018-06-29 北京市商汤科技开发有限公司 Target object detection method, device, electronic equipment and storage medium
CN108229495B (en) * 2017-06-23 2020-07-17 北京市商汤科技开发有限公司 Target object detection method and device, electronic equipment and storage medium
CN108388874A (en) * 2018-03-05 2018-08-10 厦门大学 Prawn morphological parameters method for automatic measurement based on image recognition and cascade classifier
CN108388874B (en) * 2018-03-05 2020-03-31 厦门大学 Automatic measurement method of shrimp morphological parameters based on image recognition and cascade classifier
CN108830210A (en) * 2018-06-11 2018-11-16 广东美的制冷设备有限公司 Human body detecting method and device based on image
CN108830210B (en) * 2018-06-11 2021-04-20 广东美的制冷设备有限公司 Human body detection method and device based on image
CN109977965A (en) * 2019-02-28 2019-07-05 北方工业大学 A method and device for determining detection targets in remote sensing airport images

Similar Documents

Publication Publication Date Title
CN103186790A (en) Object detecting system and object detecting method
Lee et al. Adaboost for text detection in natural scene
Pan et al. A robust system to detect and localize texts in natural scene images
CN102682287B (en) Pedestrian Detection Method Based on Saliency Information
CN101673338B (en) Fuzzy license plate identification method based on multi-angle projection
Antonacopoulos et al. ICDAR2015 competition on recognition of documents with complex layouts-RDCL2015
CN103473571B (en) Human detection method
Gerónimo et al. 2D–3D-based on-board pedestrian detection system
WO2018072233A1 (en) Method and system for vehicle tag detection and recognition based on selective search algorithm
CN103824090B (en) Adaptive face low-level feature selection method and face attribute recognition method
CN105261109A (en) Identification method of prefix letter of banknote
CN102722712A (en) Multiple-scale high-resolution image object detection method based on continuity
CN103136504A (en) Face recognition method and device
CN103824052A (en) Multilevel semantic feature-based face feature extraction method and recognition method
CN104021394A (en) Insulator image recognition method based on Adaboost algorithm
CN106156777A (en) Textual image detection method and device
CN104299009A (en) Plate number character recognition method based on multi-feature fusion
Yan et al. Chinese text location under complex background using Gabor filter and SVM
Redondo-Cabrera et al. All together now: Simultaneous object detection and continuous pose estimation using a hough forest with probabilistic locally enhanced voting
CN102147867A (en) Method for identifying traditional Chinese painting images and calligraphy images based on subject
CN101470802A (en) Object detection apparatus and method thereof
Zhou et al. Histograms of categorized shapes for 3D ear detection
Kobchaisawat et al. Thai text localization in natural scene images using convolutional neural network
Yanagisawa et al. Face detection for comic images with deformable part model
CN103984965A (en) Pedestrian detection method based on multi-resolution character association

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
AD01 Patent right deemed abandoned
AD01 Patent right deemed abandoned

Effective date of abandoning: 20181204