CN103530612A

CN103530612A - Rapid target detection method based on small quantity of samples

Info

Publication number: CN103530612A
Application number: CN201310479987.5A
Authority: CN
Inventors: 叶茂; 徐培; 占伟鹏; 黄仁杰; 张之曦
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2013-10-14
Filing date: 2013-10-14
Publication date: 2014-01-22
Anticipated expiration: 2033-10-14
Also published as: CN103530612B

Abstract

The invention discloses a rapid target detection method based on a small quantity of samples. The method mainly comprises the following steps: constructing a voting space relevant to a target by using the HOG (Histogram of Oriented Gradients) features of a small quantity of samples, calculating the HOG feature of an inquiry image and the HOG features of a detection image at different sizes, calculating the HOG distance between the inquiry image and a detection image block by adopting a window sliding way so as to locate the target in the detection image; accurately extracting the target located in the detection image in the previous step via a mean-shift algorithm, and fusing overlapped detection frames. The method is simple in process; meanwhile, detection can be performed on the basis of less samples, and the accuracy of an obtained result is higher.

Description

Fast target detection method based on a small amount of sample

Technical field

The invention belongs to computer vision and fuzzy recognition technology field, specifically a kind of design that is applied in the fast target detection method based on a small amount of sample in picture library and video.

Background technology

High speed development along with electronic information technology, the application technology of image and video has been deep into each corner of people's live and work, for example retrieval of the authentication based on image information, safety monitoring and image information data etc., in such application scenarios, conventionally need to carry out detection and Identification to a certain target appearing in image or video.Because the target in image or video often exists deformation of noise, illumination, so be still a difficult problem to the detection of a certain target in image or video.Conventionally, people can select a kind ofly based on great amount of samples, to come the mode of training classifier to carry out target detection, such scheme often need to be collected thousands of sample for a certain target, and need long time to carry out training classifier, once the target that will detect change, needs again again to collect sample training.In many application scenarioss, collecting thousands of training sample is impossible realize, for example: in the image information verification process of airport security, a passenger often only has two or three photo and me to carry out authentication, just cannot obtain about passenger's thousands of pictures and verify in the case, but the technology that need to verify from several photos; And for example, in shopping at network, in client's hand, there are several own interested commodity pictures can not to be again trade name, so need to search out these commodity on website by the mode of picture searching, now just can only just search corresponding goods picture by these samples; In addition, in safety defense monitoring system, because the target occurring is of all kinds, if carry out training classifier by collecting thousands of sample, will be a very great job, therefore, can select to be undertaken by the scheme of a small amount of sample.In addition, while carrying out target detection in image or video, also need to consider speed issue, localizing objects rapidly in image and video, if a searching algorithm is too slow, can have influence on applying of relevant art.

In prior art, " the SAR image object detection method based on Primal Sketch algorithm " of people's inventions such as the Liu Fang of Xian Electronics Science and Technology University, number of patent application is: 2011101028551, this cannot realize mainly for object detection method the shortcoming that dissimilar made Target is detected, and its implementation procedure is: 1) to former SAR image, use Primal Sketch algorithm to obtain the line segment aggregate of image result information; 2) define and calculate regularity and the regular ratio of all line segments in line segment aggregate; 3) be identified for the seed line segment aggregate appreciating in region; 4) take seed line segment as benchmark carries out region growing, obtain the candidate target region set that comprises made Target and natural target; 5) according to the Feature Selection of line segment in candidate target region, obtain final made Target.First this invention calculates SAR image, then adopts Primal Sketch algorithm to represent the line segment aggregate in SAR image, thereby generates seed line segment region by calculating line segment regularity and regular ratio.But there is following shortcoming in the method: 1) feature of the method is the line segment based on image, and such feature representation is easily subject to the impact of illumination, contrast, and there is no local feature information; 2) travelling speed is slower, is difficult to meet network picture and video frequency searching.

Summary of the invention

Technical matters to be solved by this invention is to propose a kind of fast target detection method based on a small amount of sample for the shortcoming existing in target recognition detection method in above-mentioned prior art.

The technical scheme that the present invention solves its technical matters employing is: the fast target detection method based on a small amount of sample, specifically comprises:

S1, collect the sample of the target about detecting, and form sample set R;

S2, the sample scaling of collecting in step S1 is big or small to unification, calculate the histogram of gradients HOG feature of each samples pictures in sample set R, to the samples pictures in sample set R, adopt the mode of combination of two to calculate respectively the HOG histogram distance of two each pixels of samples pictures;

S3, for size, at each location of pixels, try to achieve average μ and the variance Σ of the HOG histogram distance in described step S2;

S4, select a samples pictures as inquiry picture Q at random from sample set R, calculate its HOG feature; Test picture T for input carries out scaling on a plurality of yardsticks, and on each yardstick, calculates its HOG feature respectively;

S5, by the HOG feature of inquiry picture Q, the test picture T under a plurality of yardsticks is calculated to similarity in the mode of moving window, when similarity is greater than setting threshold, think to detect target;

S6, merge the overlapping detection target having calculated as final testing result.

Further, described HOG histogram apart from computing formula is:

wherein said h ₁and h ₂be respectively two histogram vectors,

with

represent respectively histogram vectors h ₁and h ₂i element.

Further, described step S2 specifically comprises:

S21, the mode denoising to each samples pictures employing medium filtering;

The HOG feature of each samples pictures after S22, calculation procedure S21 denoising, the corresponding histogram vectors of each location of pixels of each samples pictures after calculating;

S23, the HOG histogram distance that the samples pictures of combination of two is calculated respectively in each pixel.

Further, in described step, in S2, two samples pictures are all three-dimensional matrices after calculating HOG feature, and its HOG feature is designated as m * n * p, wherein m and n represent respectively length in the plane and the width of HOG feature, and p represents the length of each pixel position histogram vectors.

Further, described step S4 specifically comprises:

S41, inquiry picture Q is adopted to the denoising of medium filtering mode, calculate its HOG feature;

Scaling is carried out in S42, the test picture T employing medium filtering mode denoising to input on a plurality of yardsticks, calculates respectively its HOG feature on each yardstick.

Further, described step S5 comprises:

S51, adopt the mode of moving window to calculate the image block T of formed objects in the test picture T under inquiry picture Q and onesize a certain yardstick _ihOG histogram distance;

If S52 inquiry picture Q and image block T _ihOG histogram distance and the absolute value of the difference of the average μ of the HOG histogram distance of the respective pixel position described in step S3 of respective pixel is less than the variance Σ of the HOG histogram distance of respective pixel position, has a similarity to be accumulated to T _ion this image block, cumulative T _ithe similarity of upper all positions just becomes image block T _iwith the similarity of inquiry picture Q integral body, as image block T _iwhen being greater than setting threshold, whole similarity thinks to detect target.

Further, in described step S5, inquiry picture Q and image block Tibetween HOG histogram distance be:<maths TranNum="97" num="0001"><![CDATA[<math> <mrow> <msup> <mi>D</mi> <mi>i</mi> </msup> <mfenced open='[' close=']'> <mtable> <mtr> <mtd> <msub> <mi>d</mi> <mn>11</mn> </msub> </mtd> <mtd> <msub> <mi>d</mi> <mn>12</mn> </msub> </mtd> <mtd> <mo>·</mo> <mo>·</mo> <mo>·</mo> </mtd> <mtd> <msub> <mi>d</mi> <mrow> <mn>1</mn> <mi>n</mi> </mrow> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>d</mi> <mn>21</mn> </msub> </mtd> <mtd> <msub> <mi>d</mi> <mn>22</mn> </msub> </mtd> <mtd> <mo>·</mo> <mo>·</mo> <mo>·</mo> </mtd> <mtd> <msub> <mi>d</mi> <mrow> <mn>2</mn> <mi>n</mi> </mrow> </msub> </mtd> </mtr> <mtr> <mtd> <mo>·</mo> </mtd> <mtd> <mo>·</mo> </mtd> <mtd> </mtd> <mtd> <mo>·</mo> </mtd> </mtr> <mtr> <mtd> <mo>·</mo> </mtd> <mtd> <mo>·</mo> </mtd> <mtd> <msub> <mi>d</mi> <mi>jk</mi> </msub> </mtd> <mtd> <mo>·</mo> </mtd> </mtr> <mtr> <mtd> <mo>·</mo> </mtd> <mtd> <mo>·</mo> </mtd> <mtd> </mtd> <mtd> <mo>·</mo> </mtd> </mtr> <mtr> <mtd> <msub> <mi>d</mi> <mrow> <mi>m</mi> <mn>1</mn> </mrow> </msub> </mtd> <mtd> <msub> <mi>d</mi> <mrow> <mi>m</mi> <mn>2</mn> </mrow> </msub> </mtd> <mtd> <mo>·</mo> <mo>·</mo> <mo>·</mo> </mtd> <mtd> <msub> <mi>d</mi> <mi>mn</mi> </msub> </mtd> </mtr> </mtable> </mfenced> </mrow></math>]]></maths>Wherein, j<m, k<n, djkfor inquiry picture Q and image block Tibetween HOG histogram distance between the j * k histogram;

Inquiry picture Q and image block T _ibetween calculating formula of similarity be:

μ wherein _jkrepresent the j * k element in μ, Σ _jkrepresent the j * k element in Σ, function

be defined as:

n (d wherein _jk, μ _jk, Σ _jk) be gauss of distribution function, the average of this function is μ _jk, variance is Σ _jk.

Further, in described step S6, adopt means-shift algorithm to merge the overlapping detection target having calculated.

Beneficial effect of the present invention: the fast target detection method that the present invention is based on a small amount of sample is constructed the voting space about target by the histogram of gradients HOG feature of the sample with a small amount of, and the HOG feature of calculating detected image under the HOG feature of query image and different scale, adopt the mode of moving window to calculate the histogram of gradients distance of query image and detected image piece, thereby the target in detecting picture is positioned; Finally by mean-shift algorithm, the target in the detected image navigating in previous step is accurately extracted, thereby the detection block overlaping is merged mutually, its procedure is simple, can detect by the sample based on still less simultaneously, and the result degree of accuracy drawing is higher.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of the fast target detection method based on a small amount of sample of the embodiment of the present invention;

Fig. 2 is for take the testing process block diagram that people's face is example;

The schematic diagram of Fig. 3 for adopting the mode of moving window to carry out the calculating of similarity;

Fig. 4 is for take the testing result schematic diagram that people's face is example;

Fig. 5 is for take the testing result schematic diagram that automobile is example;

Fig. 6 is for take testing result schematic diagram heart-shaped and that flower is example;

Fig. 7 is the testing result schematic diagram with the artificial row of walking.

Embodiment

Below in conjunction with accompanying drawing and specific embodiment, the invention will be further elaborated.

Be illustrated in figure 1 the FB(flow block) of the fast target detection method based on a small amount of sample of the embodiment of the present invention, it specifically comprises,

S1, collect a small amount of sample about the target that will detect, and form sample set R; Wherein sample set R should be from open in industry and generally acknowledged data set or from the picture of network search engines, and allows certain deformation;

S6, merge the overlapping detection target having calculated as final testing result; Wherein, can adopt means-shift algorithm to merge overlapping to detecting target.

Wherein, described HOG histogram apart from computing formula is:

wherein said h ₁and h ₂be respectively two histogram vectors,

with

represent respectively histogram vectors h ₁and h ₂get respectively i element.

Described step S2 specifically comprises:

S21, the mode denoising to each samples pictures employing medium filtering;

S22, each samples pictures is calculated to HOG feature, the corresponding histogram vectors of each location of pixels of each samples pictures after calculating;

After two samples pictures are calculated HOG feature, are all three-dimensional matrices, its HOG feature is designated as m * n * p, and wherein m and n represent respectively length in the plane and the width of HOG feature, and p represents the length of each pixel position histogram vectors.In like manner, the inquiry picture Q relating in following step and image block T _ibetween in the calculating of similarity, also need to use above-mentioned three-dimensional matrice.

Described step S4 specifically comprises:

Described step S5 comprises:

Wherein, in described step S5, calculate query graph sheet Q and image block T _isimilarity time, need to use average μ and variance Σ in described step S3, wherein average μ and variance Σ are respectively that size is the matrix of m * n,

Inquiry picture Q and image block Tibetween HOG histogram distance be:<img TranNum="162" file="BDA0000395601840000051.GIF" he="321" img-content="drawing" img-format="GIF" inline="yes" orientation="portrait" wi="579"/>wherein, j<m, k<n, djkfor inquiry picture Q and image block Tibetween HOG histogram distance between the j * k histogram;

Inquiry picture Q and image block T _ibetween calculating formula of similarity be: μ wherein _jkrepresent the j * k element in μ, Σ _jkrepresent the j * k element in Σ, function be defined as:

Described step S6 is specially: the coordinate (x, y) of the detection target location that described step S5 examination is gone out and each similarity value are updated in means-shift algorithm, and the iteration by limited step is fused into one a plurality of target frames in same target.

For those skilled in the art can understand and implement the present invention is based on the fast target detection method of a small amount of sample, in connection with concrete embodiment, the present invention program is described in detail:

We take that people's face detects is example, be illustrated in figure 2 the FB(flow block) of testing process, the people's face samples pictures wherein comprising only has 4, the size of samples pictures is 80 * 80 pixels, each samples pictures is carried out to the HOG feature that each sample is calculated in medium filtering denoising afterwards, and each size having obtained after having calculated is 10 * 10 * 32; The HOG feature that is 10 * 10 * 32 to the size of 4 samples pictures respectively combination of two is calculated HOG histogram distance, concerning two samples, what account form adopted is the histogram difference compute histograms distance to 10 * 10 32 dimensions, calculating obtains the matrix that a size is 10 * 10 after completing, each element representation of matrix be histogram distance.4 people's face samples, combination of two, can access individual such 10 * 10 matrix.

By 4 people's face samples, obtain the matrix of 6 10 * 10, be designated as respectively D ¹, D ²..., D ⁶, the Mean Matrix μ of histogram distance is:

μ_{i, j} = \frac{D_{i, j}^{1} + D_{i, j}^{2} + \cdot \cdot \cdot + D_{i, j}^{6}}{6} i, j = 1,2, . . . 10,

μ _i,jthe capable j column element of i that represents Mean Matrix μ,

the capable j column element of i that represents k matrix.Variance matrix Σ is calculated as:

Σ_{i, j} = \frac{Σ_{k = 1}^{6} {(D_{i, j}^{k} - μ_{i, j})}^{2}}{6} i, j = 1,2, . . . 10 .

When carrying out target detection, from existing four samples, select at random people's face picture as the inquiry picture sliding in test picture, and calculate its HOG feature, be designated as Q.

Picture to be detected to input, first adopts medium filtering, then carries out multiple dimensioned scaling, then on each yardstick, calculates HOG feature.On each yardstick of the picture to be detected of inputting, adopt the mode of moving window to carry out the calculating of similarity, as shown in Figure 3, the size of window is identical with the size of inquiry picture Q.Each window all calculates a histogram similarity with inquiry picture Q, if histogram similarity is greater than the threshold value of a certain setting, thinks in window and comprises people's face.To some video in window T _i, T _icorresponding HOG feature is H ⁱ, the HOG feature that Q is corresponding is H, H ⁱwith H are all matrixes of 10 * 10 * 32, T _iand the histogram D between Q ⁱthe matrix of 10 * 10, Q and T _ibetween calculating formula of similarity be:

D wherein _jkd ⁱthe j * k element.

Because testing process is by pixel moving window in test picture, so the same people's face at test picture often all can occur that similarity is greater than the situation of predefined threshold value around, therefore after moving window completes, we are by these a plurality of position candidate in same target of mean-shift algorithm fusion, the input of mean-shift algorithm is the image coordinate (x of these position candidate, y) and the similarity of this position, this algorithm can merge overlapped candidate frame after having moved.The common practise that wherein mean-shift algorithm is those skilled in the art is not described in detail in the present patent application scheme.

In order to further illustrate the practicality of the present patent application scheme, this programme has been enumerated in addition several different testing results that target completes by method of the present invention that detect on the basis of above-mentioned example, respectively as shown in FIG. 4,5,6, 7, the testing process that it is concrete and the testing process of above-mentioned example are similar, at this, are not repeated.

Those of ordinary skill in the art will appreciate that, embodiment described here is in order to help reader understanding's principle of the present invention, should be understood to that protection scope of the present invention is not limited to such special statement and embodiment.Those of ordinary skill in the art can make various other various concrete distortion and combinations that do not depart from essence of the present invention according to these technology enlightenments disclosed by the invention, and these distortion and combination are still in protection scope of the present invention.

Claims

1. the fast target detection method based on a small amount of sample, is characterized in that, specifically comprises:

S1, collect the sample of the target about detecting, and form sample set R;

2. the fast target detection method based on a small amount of sample as claimed in claim 1, is characterized in that, described HOG histogram apart from computing formula is:

wherein said h ₁and h ₂be respectively two histogram vectors,

with represent respectively histogram vectors h ₁and h ₂i element.

3. the fast target detection method based on a small amount of sample as claimed in claim 1, is characterized in that, described step S2 specifically comprises:

S21, the mode denoising to each samples pictures employing medium filtering;

4. the fast target detection method based on a small amount of sample as claimed in claim 3, it is characterized in that, in described step in S2, two samples pictures are all three-dimensional matrices after calculating HOG feature, its HOG feature is designated as m * n * p, wherein m and n represent respectively length in the plane and the width of HOG feature, and p represents the length of each pixel position histogram vectors.

5. the fast target detection method based on a small amount of sample as claimed in claim 1, is characterized in that, described step S4 specifically comprises:

6. the fast target detection method based on a small amount of sample as claimed in claim 1, is characterized in that, described step S5 comprises:

7. the fast target detection method based on a small amount of sample as claimed in claim 6, is characterized in that, in described step S5, and inquiry picture Q and image block Tibetween HOG histogram distance be:<maths TranNum="252" num="0001"><![CDATA[<math> <mrow> <msup> <mi>D</mi> <mi>i</mi> </msup> <mfenced open='[' close=']'> <mtable> <mtr> <mtd> <msub> <mi>d</mi> <mn>11</mn> </msub> </mtd> <mtd> <msub> <mi>d</mi> <mn>12</mn> </msub> </mtd> <mtd> <mo>·</mo> <mo>·</mo> <mo>·</mo> </mtd> <mtd> <msub> <mi>d</mi> <mrow> <mn>1</mn> <mi>n</mi> </mrow> </msub> </mtd> </mtr> <mtr> <mtd> <msub> <mi>d</mi> <mn>21</mn> </msub> </mtd> <mtd> <msub> <mi>d</mi> <mn>22</mn> </msub> </mtd> <mtd> <mo>·</mo> <mo>·</mo> <mo>·</mo> </mtd> <mtd> <msub> <mi>d</mi> <mrow> <mn>2</mn> <mi>n</mi> </mrow> </msub> </mtd> </mtr> <mtr> <mtd> <mo>·</mo> </mtd> <mtd> <mo>·</mo> </mtd> <mtd> </mtd> <mtd> <mo>·</mo> </mtd> </mtr> <mtr> <mtd> <mo>·</mo> </mtd> <mtd> <mo>·</mo> </mtd> <mtd> <msub> <mi>d</mi> <mi>jk</mi> </msub> </mtd> <mtd> <mo>·</mo> </mtd> </mtr> <mtr> <mtd> <mo>·</mo> </mtd> <mtd> <mo>·</mo> </mtd> <mtd> </mtd> <mtd> <mo>·</mo> </mtd> </mtr> <mtr> <mtd> <msub> <mi>d</mi> <mrow> <mi>m</mi> <mn>1</mn> </mrow> </msub> </mtd> <mtd> <msub> <mi>d</mi> <mrow> <mi>m</mi> <mn>2</mn> </mrow> </msub> </mtd> <mtd> <mo>·</mo> <mo>·</mo> <mo>·</mo> </mtd> <mtd> <msub> <mi>d</mi> <mi>mn</mi> </msub> </mtd> </mtr> </mtable> </mfenced> </mrow></math>]]></maths>Wherein, j<m, k<n, djkfor inquiry picture Q and image block Tibetween HOG histogram distance between the j * k histogram;

Inquiry picture Q and image block T _ibetween calculating formula of similarity be: μ wherein _jkrepresent the j * k element in μ, Σ _jkrepresent the j * k element in Σ, function

be defined as:

8. the fast target detection method based on a small amount of sample as described in claim 1 to 7 any one claim, is characterized in that, adopts means-shift algorithm to merge the overlapping detection target having calculated in described step S6.