CN103136524A - Object detecting system and method capable of restraining detection result redundancy - Google Patents

Abstract

The invention provides an object detecting system and a method capable of restraining detection result redundancy. The object detecting system comprises an image receiving unit, a feature extraction unit, a detection unit and a redundancy restraining unit, wherein the image receiving unit receives to-be-detected images, the feature extraction unit extracts image features of preset types from the to-be-detected images, the detection unit inputs the extracted image features into a mixed deformable part model and utilizes the mixed deformable part model to detect a plurality of window areas including objects from the to-be-detected images, and the redundancy restraining unit removes false window areas from the plurality of window areas according to interactive relationship among the plurality of window areas.

Description

Object detection systems and the method that can suppress the testing result redundancy

Technical field

The present invention relates to vision and area of pattern recognition.More particularly, relate to a kind of system and method that can suppress the detected object of testing result redundancy.

Background technology

Object detection is an important technology in 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.

A kind of method for checking object commonly used is the multi-slide-windows detection method.In the method, a plurality of feature passages of definition and space pyramid rank.The histogrammic pyramid of usage space calculates each feature passage.In order to accelerate detection speed, set up multistage detection.Yet in testing process, the sliding window of great majority is rejected in more forward level with lower cost.Yet even like this, it detects required assessing the cost is also very high.

Therefore, need a kind of method for checking object and system that can more effectively identify object from image.

Summary of the invention

One object of the present invention is to provide a kind of object detection systems and method.This detection system and method can utilize symmetrical color parts dual mode (SC-LBP) feature to carry out object detection.

Another object of the present invention is to provide a kind of feature deriving means and method, it can reduce the extracted amount of information effectively, can obtain more effective information simultaneously.

Another object of the present invention is to provide restraining device and the method for the testing result of a kind of deformable component model (DPM), it can reduce the redundancy of the testing result of existing DPM effectively.

Another object of the present invention is to provide a kind of object detection systems and method.This detection system and method can suppress the redundancy in the testing result of deformable component model (DPM) effectively.

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, extract the characteristics of image of predefined type from image to be detected, the characteristics of image of described predefined type comprises symmetrical color parts dual mode (SC-LBP) feature, the SC-LBP feature instantiation color characteristics relation between the pixel of position symmetry in the image; Detecting unit comes detected object from image to be detected by the characteristics of image input sorter that will extract, and wherein, trains described sorter by the characteristics of image that extracts described predefined type from training sample.

Alternatively, feature extraction unit is divided into a plurality of block of pixels with image to be detected, determine the color characteristics relation between the pixel of every pair of symmetry in each block of pixels, and according to the color characteristics relation of determining determine and each block of pixels in the corresponding bi-values of every pair of symmetrical pixels as the SC-LBP condition code.

Alternatively, the pixel of described every pair of symmetry is about the central point of block of pixels.

Alternatively, the SC-LBP of each block of pixels is characterized as the bi-values sum of every pair of symmetrical pixels in this block of pixels.

Alternatively, to close be color component between symmetrical pixels and/or the magnitude relationship of brightness for described color characteristics.

Alternatively, when the pixel block size was (2a+1) * (2b+1), the SC-LBP feature of block of pixels was represented as:

$\underset{p=0}{\overset{a-1}{\mathrm{\Σ}}}\mathrm{\η}(i_p,i_{\mathrm{pm}})2^p+\underset{l=a}{\overset{2\mathrm{ab}+a+b-1}{\mathrm{\Σ}}}\mathrm{\δ}(i_l,i_{\mathrm{lm}})2^l,$

When the pixel block size is (2a+1) * 2b, the SC-LBP feature of block of pixels is represented as:

$\underset{l=0}{\overset{2\mathrm{ab}+b-1}{\mathrm{\Σ}}}\mathrm{\δ}(i_l,i_{\mathrm{lm}})2^l,$

When the pixel block size is that 2a * (2b+1), the SC-LBP feature of block of pixels is represented as:

$\underset{p=0}{\overset{a-1}{\mathrm{\Σ}}}\mathrm{\η}(i_p,i_{\mathrm{pm}})2^p+\underset{l=a}{\overset{2\mathrm{ab}+a-1}{\mathrm{\Σ}}}\mathrm{\δ}(i_l,i_{\mathrm{lm}})2^l,$

When the pixel block size is 2a * 2b, the SC-LBP feature of block of pixels is represented as:

$\underset{l=0}{\overset{2\mathrm{ab}-1}{\mathrm{\Σ}}}\mathrm{\δ}(i_l,i_{\mathrm{lm}})2^l,$

Wherein, a, b are the integer greater than zero, i _lL pixel in the expression block of pixels, i _pP pixel in the expression block of pixels, i _cThe center of expression block of pixels, i _lmExpression i _lAbout i _cCentrosymmetric pixel, i _pmI _pAbout i _cCentrosymmetric pixel, η (i _p, i _pm) expression i _pAnd i _pmBetween brightness relationship, δ (i _l, i _lm) expression i _lAnd i _lmBetween the color component relation.

Alternatively,

Wherein, the brightness of I () expression pixel, the red component of R () expression pixel, the green component of G () expression pixel, the blue component of B () expression pixel, T represents predetermined threshold.

Alternatively, the size of predetermined threshold T has determined the inhibition degree to color noise, and wherein, T is less, and the sensitivity to noise is stronger.

Alternatively, the characteristics of image of described predefined type also comprises the gradient orientation histogram feature.

Alternatively, use deformable component model (DPM) method to train described sorter.

Alternatively, described sorter is to mix the deformable component model, and detecting unit detects by the described mixing deformable component model of characteristics of image input that will extract a plurality of window areas that are believed to comprise object from image to be detected.

Alternatively, described object detection systems also comprises: redundancy suppresses the unit, removes pseudo-window area from described a plurality of window areas according to the interactive relation between described a plurality of window areas.

Alternatively, redundancy suppresses the unit and comprises: feature information extraction unit, each the window area characteristic information extraction that detects from detecting unit; Redundancy is removed the unit, 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 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.

Alternatively, redundancy is removed the unit by maximizing following equation and judge and remove pseudo-window area:

Wherein, M is the quantity of the described a plurality of window areas of expression;

φ (x _i, y _i)=y _iX _i

x _i=(v _i(s), Z), v _i(s) PTS 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 comprises, 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 the deformable component model that i window area use; y _iWhat represent 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;

The expression model parameter, d _ijRepresent the interactive relation between i window area and j window area.

Alternatively, 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),

SS trains to obtain by the predetermined structure sorting technique

Alternatively, the interactive relation between window area comprise that root-root is mutual, root-parts mutual, at least one in mutual of parts-parts.

Alternatively, root-root has embodied the overlapping characteristic between the root in different windows zone alternately.

Alternatively, root-parts have embodied the root in different windows zone and the overlapping characteristic between parts alternately.

Alternatively, parts-parts have embodied the overlapping characteristic between the parts in different windows zone alternately.

Alternatively, root-root is represented as alternately

The matrix of K * K, the arbitrary element in this matrix

Be expressed as following equation:

Wherein, v _j(s) PTS of j window area of expression, ol (v _i(l ₀), v _j(l ₀)) Duplication between the root of the expression root of i window area and j window area, v _j(c) expression detects the index of the deformable component model that j window area use, and m, n represent the index of described arbitrary element.

Alternatively, root-parts are represented as alternately K * (matrix of K * D), arbitrary element in this matrix

Can be expressed as following equation:

Wherein, v _j(s) PTS of j window area of expression, g ∈ [1, D], D represents the quantity of parts, ol (v _i(l ₀), v _j(l _g)) Duplication between g parts of the expression root of i window area and j window area, v _j(c) expression detects the index of the deformable component model that j window area use, and m, n represent the index of described arbitrary element.

Alternatively, parts-parts are represented as alternately (K * D) * (matrix of K * D), arbitrary element in this matrix

Can be expressed as following equation:

Wherein, v _i(s _e) score of e parts of i window area of expression, v _j(s _g) score of g parts of j window area of expression, e ∈ [1, D], g ∈ [1, D], D represents the quantity of parts, ol (v _i(l _e), v _j(l _g)) Duplication between e parts of expression i window area and g parts of j window area, v _j(c) expression detects the index of the deformable component model that j window area use, and m, n represent the index of described arbitrary element.

Alternatively, y _i∈ { 0,1}, y _j∈ { 0,1}.

Alternatively, use greedy algorithm to make SS maximize.

Alternatively, described predetermined structure sorting technique is the structuring support vector machine.

Another invention of the present invention provides a kind of method for checking object, comprising: receive image to be detected; Extract the characteristics of image of predefined type from image to be detected, the characteristics of image of described predefined type comprises symmetrical color parts dual mode (SC-LBP) feature, the SC-LBP feature instantiation color characteristics relation between the pixel of position symmetry in the image; Come detected object from image to be detected by the characteristics of image input sorter that will extract, wherein, train described sorter by the characteristics of image that extracts described predefined type from training sample.

Alternatively, image to be detected is divided into a plurality of block of pixels, determine the color characteristics relation between the pixel of every pair of symmetry in each block of pixels, and according to the color characteristics relation of determining determine and each block of pixels in the corresponding bi-values of every pair of symmetrical pixels as the SC-LBP condition code.

Alternatively, the pixel of described every pair of symmetry is about the central point of block of pixels.

Alternatively, the SC-LBP of each block of pixels is characterized as the bi-values sum of every pair of symmetrical pixels in this block of pixels.

Alternatively, to close be color component between symmetrical pixels and/or the magnitude relationship of brightness for described color characteristics.

Alternatively, the size of predetermined threshold T has determined the inhibition degree to color noise, and wherein, T is less, and the sensitivity to noise is stronger.

Alternatively, the characteristics of image of described predefined type also comprises the gradient orientation histogram feature.

Alternatively, use deformable component model (DPM) method to train described sorter.

Alternatively, wherein, described sorter is to mix the deformable component model, detects by the described mixing deformable component model of characteristics of image input that will extract a plurality of window areas that are believed to comprise object from image to be detected.

Alternatively, described method for checking object also comprises: remove pseudo-window area from described a plurality of window areas according to the interactive relation between described a plurality of window areas.

Alternatively, the step of the pseudo-window area of removal comprises: from each window area characteristic information extraction of detecting unit detection; Utilize 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 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.

Alternatively, judge and remove pseudo-window area by maximizing following equation:

Wherein, M is the quantity of the described a plurality of window areas of expression;

φ (x _i, y _i)=y _iX _i

x _i=(v _i(s), Z), v _i(s) PTS 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 comprises, 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 the deformable component model that i window area use; y _iWhat represent 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;

The expression model parameter, d _ijRepresent the interactive relation between i window area and j window area.

Alternatively, 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), SS trains to obtain by the predetermined structure sorting technique

Alternatively, the interactive relation between window area comprise that root-root is mutual, root-parts mutual, at least one in mutual of parts-parts.

Alternatively, described predetermined structure sorting technique structuring support vector machine.

Another aspect of the present invention provides a kind of Characteristic of Image extracting method, comprising: image is divided into a plurality of block of pixels; Obtain the color characteristics of each pixel in each block of pixels; Determine the color characteristics relation between every pair of symmetrical pixels in each block of pixels; According to the color characteristics relation of determining determine and each block of pixels in the corresponding bi-values of every pair of symmetrical pixels as the SC-LBP condition code.

Alternatively, the pixel of described every pair of symmetry is about the central point of block of pixels.

Alternatively, the SC-LBP of each block of pixels is characterized as the bi-values sum of every pair of symmetrical pixels in this block of pixels.

Alternatively, to close be color component between symmetrical pixels and/or the magnitude relationship of brightness for described color characteristics.

Alternatively, when the pixel block size was (2a+1) * (2b+1), the SC-LBP feature of block of pixels was represented as: $\underset{p=0}{\overset{a-1}{\mathrm{\Σ}}}\mathrm{\η}(i_p,i_{\mathrm{pm}})2^p+\underset{l=a}{\overset{2\mathrm{ab}+a+b-1}{\mathrm{\Σ}}}\mathrm{\δ}(i_l,i_{\mathrm{lm}})2^l,$

When the pixel block size is (2a+1) * 2b, the SC-LBP feature of block of pixels is represented as:

$\underset{l=0}{\overset{2\mathrm{ab}+b-1}{\mathrm{\Σ}}}\mathrm{\δ}(i_l,i_{\mathrm{lm}})2^l,$

When the pixel block size is that 2a * (2b+1), the SC-LBP feature of block of pixels is represented as:

$\underset{p=0}{\overset{a-1}{\mathrm{\Σ}}}\mathrm{\η}(i_p,i_{\mathrm{pm}})2^p+\underset{l=a}{\overset{2\mathrm{ab}+a-1}{\mathrm{\Σ}}}\mathrm{\δ}(i_l,i_{\mathrm{lm}})2^l,$

When the pixel block size is 2a * 2b, the SC-LBP feature of block of pixels is represented as:

$\underset{l=0}{\overset{2\mathrm{ab}-1}{\mathrm{\Σ}}}\mathrm{\δ}(i_l,i_{\mathrm{lm}})2^l,$

Wherein, a, b are the integer greater than zero, i _lL pixel in the expression block of pixels, i _pP pixel in the expression block of pixels, i _cThe center of expression block of pixels, i _lmExpression i _lAbout i _cCentrosymmetric pixel, i _pmI _pAbout i _cCentrosymmetric pixel, η (i _p, i _pm) expression i _pAnd i _pmBetween brightness relationship, δ (i _l, i _lm) expression i _lAnd i _lmBetween the color component relation.

Alternatively, the size of predetermined threshold T has determined the inhibition degree to color noise, and wherein, T is less, and the sensitivity to noise is stronger.

Another invention of the present invention provides a kind of object detection systems, comprising: image receiving unit receives image to be detected; Feature extraction unit is from the characteristics of image of image extraction predefined type to be detected; Detecting unit, mix the deformable component model by the characteristics of image input of will extract, thereby utilize mixing deformable component model to detect a plurality of window areas that are believed to comprise object from image to be detected, redundancy suppresses the unit, removes pseudo-window area from described a plurality of window areas according to the interactive relation between described a plurality of window areas.

Alternatively, redundancy suppresses the unit and comprises: feature information extraction unit, each the window area characteristic information extraction that detects from detecting unit; Redundancy is removed the unit, 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 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.

Alternatively, redundancy is removed the unit by maximizing following equation and judge and remove pseudo-window area:

Wherein, M is the quantity of the described a plurality of window areas of expression;

φ (x _i, y _i)=y _iX _i

x _i=(v _i(s), Z), v _i(s) PTS 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 comprises, 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 the deformable component model that i window area use; y _iWhat represent 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;

The expression model parameter, d _ijRepresent the interactive relation between i window area and j window area.

Alternatively, 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),

SS trains to obtain by the predetermined structure sorting technique

Alternatively, the interactive relation between window area comprise that root-root is mutual, root-parts mutual, at least one in mutual of parts-parts.

Alternatively, root-root has embodied the overlapping characteristic between the root in different windows zone alternately.

Alternatively, root-parts have embodied the root in different windows zone and the overlapping characteristic between parts alternately.

Alternatively, parts-parts have embodied the overlapping characteristic between the parts in different windows zone alternately.

Alternatively, root-root is represented as alternately

The matrix of K * K, the arbitrary element in this matrix

Be expressed as following equation:

Wherein, v _j(s) PTS of j window area of expression, ol (v _i(l ₀), v _j(l ₀)) Duplication between the root of the expression root of i window area and j window area, v _j(c) expression detects the index of the deformable component model that j window area use, and m, n represent the index of described arbitrary element.

Alternatively, root-parts are represented as alternately

K * (matrix of K * D), arbitrary element in this matrix

Can be expressed as following equation:

Wherein, v _j(s) PTS of j window area of expression, g ∈ [1, D], D represents the quantity of parts, ol (v _i(l ₀), v _j(l _g)) Duplication between g parts of the expression root of i window area and j window area, v _j(c) expression detects the index of the deformable component model that j window area use, and m, n represent the index of described arbitrary element.

Alternatively, parts-parts are represented as alternately

(K * D) * (matrix of K * D), arbitrary element in this matrix

Can be expressed as following equation:

Wherein, v _i(s _e) score of e parts of i window area of expression, v _j(s _g) score of g parts of j window area of expression, e ∈ [1, D], g ∈ [1, D], D represents the quantity of parts, ol (v _i(l _e), v _j(l _g)) Duplication between e parts of expression i window area and g parts of j window area, v _j(c) expression detects the index of the deformable component model that j window area use, and m, n represent the index of described arbitrary element.

Alternatively, y _i∈ { 0,1}, y _j∈ { 0,1}.

Alternatively, use greedy algorithm to make SS maximize.

Alternatively, described predetermined structure sorting technique is the structuring support vector machine.

Alternatively, the characteristics of image of described predefined type comprises symmetrical color parts dual mode (SC-LBP) feature, the SC-LBP feature instantiation color characteristics relation between the pixel of position symmetry in the image.

Another aspect of the present invention provides a kind of method for checking object, comprising: receive image to be detected;

Extract the characteristics of image of predefined type from image to be detected; Mix the deformable component model by the characteristics of image input of will extract, thereby utilize mixing deformable component model to detect a plurality of window areas that are believed to comprise object from image to be detected, remove pseudo-window area according to the interactive relation between described a plurality of window areas from described a plurality of window areas.

Alternatively, the step of the pseudo-window area of removal comprises: from each window area characteristic information extraction of detecting unit detection; Utilize the characteristic information that extracts to determine described interactive relation, to remove pseudo-window area from described a plurality of window areas.

According to method for checking object of the present invention and system, can reduce the quantity of information of extraction and obtain simultaneously more effective information by effective use colouring information when extracting feature, thereby can reduce calculated amount when detecting, improve accuracy of detection and accelerate detection speed.In addition, can effectively reduce the redundancy of testing result according to method for checking object of the present invention and system.

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 illustrates the block diagram according to the object detection systems of the object in the detected image of the embodiment of the present invention;

Fig. 2 shows the diagram of the piece of the pre-sizing with different size parity;

Fig. 3 illustrates the block diagram that suppresses the unit according to the redundancy of the embodiment of the present invention;

Fig. 4 illustrates the process flow diagram of the method for checking object of the object in detected image according to an embodiment of the invention;

Fig. 5 illustrates the process flow diagram of the method for the redundancy that suppresses according to an embodiment of the invention the DPM sorter.

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 represents identical element.

Fig. 1 illustrates the block diagram according to the object detection systems 100 of the object in the detected image of the embodiment of the present invention.

Detection system 100 comprises image receiving unit 110, feature extraction unit 120, detecting unit 130.

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 various characteristics of image commonly used such as gradient orientation histogram (HOG) feature, parts dual mode (LBP) feature, trellis depth feature (GDF) and/or yardstick invariant features conversion (SIFT) feature.

Preferably, the characteristics of image of extraction comprises symmetrical color LBP (SC-LBP) feature of the present invention's proposition that will be described below at least.

Existing LBP feature has higher differentiation, and changes very stable for monochromatic gray level.Yet the LBP feature has very high dimension and can not use colouring information.And can apparent color information according to SC-LBP feature of the present invention, and can directly obtain feature consistency between block of pixels and mirror image thereof.The SC-LBP feature has embodied the color characteristics relation between the symmetrical pixels (or mirror image) of pixel in the image and this pixel by dual mode.Specifically, when extracting the SC-LBP feature, image is divided into a plurality of block of pixels, according to the color characteristics relation condition (for example, the magnitude relationship between color component, the magnitude relationship between brightness etc.) that arranges determine with each block of pixels in the corresponding bi-values of every pair of symmetrical pixels as the SC-LBP condition code.

Specifically, in order to obtain the SC-LBP feature of each block of pixels, can with this block of pixels in the corresponding bi-values addition of every pair of symmetrical pixels.

In one example, when extracting the SC-LBP feature, image is divided into the block of pixels of pre-sizing, for each block of pixels, calculates SC-LBP feature c according to following equation (1):

.....(1)

Wherein, a, b are positive integer, i _lL pixel in the expression block of pixels, i _pP pixel in the expression block of pixels, i _cThe center of expression block of pixels, i _lmExpression i _lAbout i _cCentrosymmetric pixel, i _pmI _pAbout i _cCentrosymmetric pixel.

Brightness relationship (for example, η (i between two pixels of η () expression symmetry _p, i _pm) expression i _pAnd i _pmBetween the magnitude relationship of brightness), color component relation (for example, the δ (i between symmetrical two pixels of δ () expression _l, i _lm) expression i _lAnd i _lmBetween the magnitude relationship of color component).

In example, utilize bi-values to represent above-mentioned magnitude relationship below.

Wherein, brightness (for example, the I (i of I () expression pixel _p) expression pixel i _pBrightness, I (i _pm) expression pixel i _pmBrightness), red component (for example, the R (i of R () expression pixel _p) expression pixel i _pRed component, R (i _pm) expression pixel i _pmRed component), green component (for example, the G (i of G () expression pixel _p) expression pixel i _pGreen component, G (i _pm) expression pixel i _pmGreen component), blue component (for example, the B (i of B () expression pixel _p) expression pixel i _pBlue component, B (i _pm) expression pixel i _pmBlue component), T represents for the threshold value that suppresses noise.The size of T has determined the inhibition degree to color noise.T is less, and the sensitivity to noise is stronger.

Should be appreciated that, the above calculates η (i _p, i _pm), δ (i _l, i _lm) example be only exemplary, can show brightness between pixel and/or the magnitude relationship of color component with other method for expressing.For example, value condition exchange of 1,0 etc.In addition, also can not use threshold value T.

Relevant in the parity of the size (or length) of different directions with the piece of the pre-sizing of division according to the calculating of SC-LBP condition code of the present invention, so equation (1) shows the method for calculating SC-LBP condition code c for the piece of the pre-sizing of different size parity.

Fig. 2 shows the diagram of the piece of the pre-sizing with different size parity.

In addition, should be appreciated that, can be numbered the pixel in block of pixels with any order.

By top formula as can be known, for the block of pixels of one (2a+1) * (2b+1) size, to be a scale-of-two length be the (numerical value of 2 * a * b+a+b) to its condition code; For the block of pixels of (2a+1) * 2b size, to be a scale-of-two length be the (numerical value of 2 * a * b+b) to its condition code; For the block of pixels of the size of a 2a * (2b+1), to be a scale-of-two length be the (numerical value of 2 * a * b+a) to its condition code; For the block of pixels of a 2a * 2b size, to be a scale-of-two length be the (numerical value of 2 * a * b) to its condition code.

For example, when the block of pixels of dividing was 3 * 3 block of pixels, SC-LBP feature c can be expressed as followsin:

c＝η(i ₀，i ₄)2 ⁰+δ(i ₁，i ₅)2 ¹+δ(i ₂，i ₆)2 ²+δ(i ₃，i ₇)2 ³。

The above shows the preferred computing method of SC-LBP condition code, wherein, and η (i _p, i _pm) and δ (i _l, i _lm) be used for the color characteristics relation condition of the bi-values of expression SC-LBP condition code.Yet, the invention is not restricted to this, the bi-values that other different color characteristics relation conditions calculate every pair of symmetrical pixels also can be set obtain the SC-LBP feature, for example, can not use monochrome information or color component information, can be set to 1 when color characteristics information is consistent, other situations are set to zero etc.In addition, should be appreciated that, bi-values of the present invention is not limited to 1 and 0, can be also other two values.

In addition, although adopted central point in the example that illustrates in the above, however symmetrical other position symmetric forms that wait that can be also line of the symmetric relation between the every pair of symmetrical pixels.Yet, preferably adopt central point, can obtain more quantity of information like this.

According to SC-LBP characteristic use of the present invention color and the symmetry information of image, and has following characteristic: have fixing mapping relations between the condition code of the block of pixels of the condition code of a block of pixels and its mirror image symmetry, and this mirror symmetry is conducive to the detection and Identification of object.

In a preferred embodiment of the invention, extract simultaneously SC-LBP feature and HOG feature.

Detecting unit 130 utilizes the Image Feature Detection object that extracts.Detecting unit 130 can come detected object with the sorter of training in advance.Can extract described characteristics of image (for example, HOG feature and SC-LBP feature, LBP feature etc.) the positive sample of generation and negative sample and come training classifier from training sample image.

For example, positive sample corresponding to training sample in the corresponding characteristics of image of object of mark, negative sample corresponding to training sample in the corresponding characteristics of image of background.

Can adopt existing various sorter training method to come training classifier, for example, can use the training methods such as support vector machine (SVM), Boost algorithm, deformable component model (DPM) to come training classifier.

Extract the SC-LBP feature in the situation that the characteristics of image that extracts comprises, can reduce the characteristics of image dimension of extraction, thereby no matter when training classifier or when detected object from image to be detected, can reduce calculated amount and raising speed.

Because the technology of utilizing various known training methods to train after extracting feature is known, will be not described in detail at this.

In one embodiment of the invention, use is based on the sorter of deformable component model (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 based on the score in the sorter of DPM:

$f(p_0,...p_n)=\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{F}_i\·\mathrm{\φ}(H,p_i)-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{d}_i\·{\mathrm{\φ}}_d(d{x}_i,d{y}_i)+b,---\left(2\right)$

Wherein,

The data item in the parts definition of the root of object and object, $\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{d}_i\·{\mathrm{\φ}}_d({\mathrm{dx}}_i,{\mathrm{dy}}_i)$ 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, F _iThat (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) represent by the characteristics of image pyramid at p _iThe feature (for example, SC-LBP feature, HOG feature) of extracting, 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 (2) as scoring function, and it depends on the concrete DPM model of use.

Make z=(p ₀... p _n), equation 2 can be written as:

f(z)＝β·Ψ(H，z)， (3)

Wherein, 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, by utilizing positive negative sample, this model training is obtained parameter F _i, d _iAnd b.

By the characteristics of image pyramid at p _iWhen extracting feature, need at first to set up image pyramid for the image that is extracted feature (for example, image to be detected, training sample image).The pyramidal bottom is the image of original resolution, and along with arriving pyramidal top layer from pyramidal lowermost layer, image is reduced image to an intended size gradually.Utilize a plurality of (for example, K) window of pre-sizing comes the scan image pyramid, the corresponding concrete DPM model of the window of each pre-sizing, there be like this K DPM model, this K one of DPM model-composing mixes the DPM model, and each DPM model is as mixing DPM model one-component.In each window (for example, size is the window of 8 * 8 prearranged multiple), for piece (for example, 8 * 8 the piece) calculated characteristics (for example, HOG feature, SC-LBP feature) of pre-sizing wherein.

When extracting feature in training process, usually there are two kinds of feature extraction modes.A kind of mode is to annotate the root of object and the position of parts in positive negative sample acceptance of the bid, thereby utilizes image pyramid directly to extract corresponding feature at root and the parts of mark according to the said extracted mode, and does not need to scan whole image pyramid.Another kind of mode is only to mark the position of the root of object in positive negative sample, and does not mark the position of the parts of object.Dynamically search thereby need to scan according to the said extracted mode so whole image pyramid in training process the component locations that makes the score maximum, to extract feature in this position.This can realize by many examples training algorithm or the algorithm of support vector machine that contains hidden variable.Because these algorithms are that algorithm commonly used in DPM will no longer describe in detail.

Detecting unit 130 extracts feature as the input of DPM model according to the said extracted mode in testing process.When the score of DPM model during greater than predetermined threshold, think to have object in this window area; Otherwise, think not have object in this window area.

As mentioned above, when the sorter that trains in utilization detects, use predetermined window to treat detected image with different scale in all positions and scan, thereby obtain a plurality of window areas.During less than predetermined threshold, this window area represents not exist in this window object for negative window area when the classification score of window area; During more than or equal to predetermined threshold, this window area is positive window area, is illustrated in this window and has object when the classification score of window area.Therefore, may obtain a plurality of overlapping positive window areas when utilizing sorter to detect, thereby obtain a plurality of testing results.In order to reduce redundancy, reduce the quantity of overlapping testing result, usually use non-maximum (Non-maximum suppression) technology that suppresses.Yet in some specific occasions, object is blocked/combines with other objects and object space layout and overlapping complicacy, so this technology can cause producing the positive window area of a lot of puppets, has greatly reduced accuracy of detection.

In order to address this problem, the present invention proposes a kind of Redundancy-Restraining Technique.According to another embodiment of the present invention, object detection systems 100 comprises that also redundancy suppresses the unit (not shown).Redundancy suppresses the unit can remove pseudo-window area from the window area of detecting unit 130 outputs.

Fig. 3 illustrates the block diagram that suppresses the unit according to the redundancy of the embodiment of the present invention.

Redundancy suppresses the unit and comprises that feature information extraction unit 141, redundancy remove unit 142.

Feature information extraction unit 141 is from each window area characteristic information extraction of detecting unit 130 outputs.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 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 _l，..dd _D，l ₀，l _l...l _D，c)， (4)

Wherein, l ₀Position and the yardstick of the root of object, l _l... l _DPosition and the yardstick of the parts of object, the quantity of the parts of D indicated object, s ₀The score of root, s _l... s _DThe score of parts, dd _l..dd _DIt is the distortion cost of parts, b is the skew that the DPM model in the mixing DPM model that uses causes, s is the PTS of window area, c be object component index (namely, represent that i window area obtained by c the DPM model detection that mixes in the DPM model), 1≤c≤K, K represent to mix the quantity of the component of DPM.

Should be appreciated that, although top v _iComprise much information, but should be appreciated that, can only extract as required partial information wherein.

Redundancy is removed unit 142 and is utilized the characteristic information that extracts to determine from the interactive relation between the window area of detecting unit 130 outputs, to remove pseudo-window area from the window area of detecting unit 130 outputs.This interactive relation has embodied the overlapping characteristic between window area.

Specifically, suppose x _iBe the characteristic information that extracts from window area i, whole 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 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).Make Y={y _i: i=1...M}, use must being divided into of vectorial Y mark image X:

Wherein, φ (x _i, y _i)=y _iX _i,

The expression 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 j window area.

Can use the φ (x of precognition _i, y _i),

SS trains to obtain by existing structuring sorting technique (for example, structuring SVM algorithm, Boost algorithm etc.)

Preferably, use structuring SVM algorithm.Owing to obtaining with the structuring sorting technique

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 are 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 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

With

For example, when using all above-mentioned three kinds when mutual,

Represent the interactive relation between the root between i window area and j window area.Be appreciated that

It is the matrix of K * K.In one example, the arbitrary element in this matrix

(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 (6):

Here, ol (v _i(l ₀), v _j(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 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 j window area.Be appreciated that

K * (matrix of K * D), arbitrary element in this matrix

Can be expressed as following equation (7):

Here, g ∈ [1, D], ol (v _i(l ₀), v _j(l _g)) Duplication between the root of expression i window area and g parts of j window area.

Represent parts between i window area and j window area and the interactive relation between parts.Be appreciated that

It is (K * D) * (matrix of K * D).In one example, the arbitrary element in this matrix

Can be expressed as following equation (9):

....(9)

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 parts of j window area.

Calculating makes equation 5 be maximum Y, and this calculating can be represented as argmax _Y S (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 5 for maximum Y, for example, can use the mode of enumerating.

In another embodiment of the present invention, calculate with greedy algorithm and make equation 5 be maximum Y.

Fig. 4 illustrates the process flow diagram of the method for checking object of the object in detected image according to an embodiment of the invention.

In step 401, receive image to be detected.

In step 402, extract the characteristics of image of predefined type from image to be detected.The characteristics of image that extracts can be various characteristics of image commonly used such as gradient orientation histogram (HOG) feature, parts dual mode (LBP) feature, trellis depth feature (GDF) and/or yardstick invariant features conversion (SIFT) feature.Preferably, the characteristics of image of extraction comprises above-described according to symmetrical color LBP of the present invention (SC-LBP) feature at least.

In step 403, the sorter of inputting training in advance by the characteristics of image that will extract comes detected object from image to be detected.Can train described sorter by the characteristics of image that extracts described predefined type from training sample.

Can use the training methods such as support vector machine (SVM), Boost algorithm, deformable component model (DPM) to come training classifier.Preferably, come training classifier to obtain mixing the deformable component model with DPM.

Owing to having pseudo-result in the result that obtains when the sorter that uses the DPM training detects, therefore need to carry out redundancy and suppress.

Fig. 5 illustrates the process flow diagram of the method for the redundancy that suppresses according to an embodiment of the invention the DPM sorter.

In step 501, from each the window area characteristic information extraction that is believed to comprise object that detects in step 403.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 502, utilize the characteristic information that extracts to remove pseudo-window area from described a plurality of window areas.Specifically, maximize the result of formula (5), thereby the window area with binary score of the pseudo-window area of sign is judged as pseudo-window area.

According to method for checking object of the present invention and system, can reduce the quantity of information of extraction and obtain simultaneously more effective information by effective use colouring information when extracting feature, thereby can reduce calculated amount when detecting, improve accuracy of detection and accelerate detection speed.In addition, can effectively reduce the redundancy of testing result according to method for checking object of the present invention and system.

Although specifically shown with reference to its exemplary embodiment and described the present invention, but it should be appreciated by those skilled in the art, in the situation that do not break away from the spirit and scope of the present invention that claim limits, can carry out various changes on form and details to it.

Claims (9)

1. object detection systems that can suppress the testing result redundancy comprises:

Image receiving unit receives image to be detected;

Feature extraction unit is extracted characteristics of image from image to be detected;

Detecting unit mixes the deformable component model by the characteristics of image input that will extract, thereby utilizes mixing deformable component model to detect a plurality of window areas that are believed to comprise object from image to be detected,

Redundancy suppresses the unit, removes pseudo-window area from described a plurality of window areas according to the interactive relation between described a plurality of window areas.

2. object detection systems according to claim 1, wherein, redundancy suppresses the unit and comprises:

The feature information extraction unit is from each window area characteristic information extraction of detecting unit detection;

Redundancy is removed the unit, utilizes the characteristic information that extracts to determine described interactive relation, to remove pseudo-window area from described a plurality of window areas.

3. object detection systems according to claim 2, wherein, described characteristic information comprises 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.

4. object detection systems according to claim 2, wherein, redundancy is removed the unit by maximizing following equation and judge and remove pseudo-window area:

Wherein, M is the quantity of the described a plurality of window areas of expression;

φ (x _i, y _i)=y _iX _i x _i=(v _i(s), Z), v _i(s) PTS 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 comprises, 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 the deformable component model that i window area use; y _iWhat represent 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;

The expression model parameter, d _ijRepresent the interactive relation between i window area and j window area.

5. object detection systems according to claim 4, wherein, when described equation maximized, the window area with binary score of the pseudo-window area of sign was judged as pseudo-window area.

6. object detection systems according to claim 4, wherein, use the φ (x of precognition _i, y _i),

SS trains to obtain by the predetermined structure sorting technique

7. object detection systems according to claim 1, wherein, the interactive relation between window area comprises that root-root is mutual, root-parts are mutual, at least one in mutual of parts-parts.

8. object detection systems according to claim 1, wherein, described characteristics of image comprises symmetrical color parts dual mode (SC-LBP) feature, the SC-LBP feature instantiation color characteristics relation between the pixel of position symmetry in the image.

9. method for checking object that can suppress the testing result redundancy comprises:

Receive image to be detected;

Extract characteristics of image from image to be detected;

Mix the deformable component model by the characteristics of image input that will extract, thereby utilize mixing deformable component model to detect a plurality of window areas that are believed to comprise object from image to be detected,

Remove pseudo-window area from described a plurality of window areas according to the interactive relation between described a plurality of window areas.

Images