CN101655910A - Training system, training method and detection method - Google Patents
Training system, training method and detection method Download PDFInfo
- Publication number
- CN101655910A CN101655910A CN200810210129A CN200810210129A CN101655910A CN 101655910 A CN101655910 A CN 101655910A CN 200810210129 A CN200810210129 A CN 200810210129A CN 200810210129 A CN200810210129 A CN 200810210129A CN 101655910 A CN101655910 A CN 101655910A
- Authority
- CN
- China
- Prior art keywords
- feature
- training
- unique point
- image
- zone
- 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
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a training system, a training method and a detection method, which are used in a training system of a classifier which is used for differentiating targets and backgrounds and isacquired through training in images or videos. The training system comprises an acquiring unit, a feature extracting unit and a training unit, wherein the acquiring unit is used for acquiring targetpictures and background pictures as a training picture set; the feature extracting unit is used for serving pixels with preset image features as feature points in the training picture set and extracting a distribution feature set of the feature points; the training unit is used for training the distribution feature set so as to acquire the classifier; and the distribution feature set at least comprises a directional entropy feature. Therefore, the training system, the training method and the detection method can improve the accuracy of target detection.
Description
Technical field
The existence that the present invention relates to recognition target image in image or video data whether with training system, training method and the detection method of the position that exists.
Background technology
In the prior art, in video or image, detect target two class methods are arranged, the first kind is to adopt the still image feature to set up the sorter of distinguishing target and background, detects target with this sorter in image, then each frame is considered as piece image for video and detects.Second class is the frame-to-frame correlation in conjunction with static nature and video, and information such as motion, sound detect objects in video.Wherein, the method for still image is the basis of detecting.
At present, adopt class Ha Er (Haar-like) rectangular characteristic to detect target in the still image, select the feature of employing automatically with the method for boost (method of machine learning), but rectangular characteristic is not very high to correct rate of target detection.For the pedestrian in the video, because people's motion has unique feature, from the difference diagram of frame-to-frame differences component and distortion, can extract feature, obtain sorter thereby train, but this can not be used for the situation of camera motion with static nature about the travel direction amplitude.In addition, in the prior art, also there is following situation: the rectangular characteristic of still image is promoted, add to tilt the features such as polygon of 45 degree, but the feature of these classes Ha Er is the same with rectangular characteristic, does not have the specific aim to target.
In addition, pedestrian in the feature detection image of employing directivity histogram of gradients (HoG), gradient is all asked in each position of target, as feature, adopt support vector machine (SVM) to train the summation of the gradient of all directions and the gradient summation ratio between the zone asked for.Because histogram has the meaning of statistics, can adapt to target among a small circle with angle in variation, but attitude when changing greatly verification and measurement ratio still remain raising.The technical scheme that detects in conjunction with static directivity histogram of gradients in addition, is also arranged by the optical flow field of video is got the motion feature that the directivity histogram feature obtains the pedestrian.The directivity histogram of gradients also is based on the feature of rectangular block, and the summation of feature in the statistics block, and the characteristic allocation ratio between the computing block are not equally considered the distribution situation of feature in the piece yet.
The verification and measurement ratio of target and still image algorithm height correlation in the video, the feature extraction in the image is the basis of target detection.
Usually become zone (mostly being rectangle or parallelogram) to ask for the statistical value of characteristics of image in the zone image division.But Haar-like feature and HoG feature etc. have all only adopted the summation or the weighted sum of regional interior all eigenwerts, the not distribution situation of consideration of regional inter characteristic points.As shown in Figure 6, the all directions gradient summation of inside, rectangular area may be similar in Target Photo (is example with the human body) and the background picture, the HoG feature can't be distinguished, but being distributed in of marginal point has systematicness in the Target Photo, and the distribution in background picture is comparatively mixed and disorderly.Especially in the human body contour outline part, the marginal point that gradient direction is identical often forms the lines with the gradient direction quadrature, thereby along the gradient direction rotatable coordinate axis, the ordinate of these marginal points may concentrate on a minizone.
Therefore, in view of the above problems, inventor moving party tropism entropy feature, it is used for the regular degree that characteristic feature point distributes, and this feature combines with other statistical natures based on the zone, thereby can effectively improve verification and measurement ratio, reduces the false drop rate of background.
Summary of the invention
In view of the above problems, the object of the invention is to provide training system, training method and the detection method that can improve correct rate of target detection.
According to an aspect of the present invention, provide to be used at image or video by training the training system of the sorter that obtains to distinguish target and background, it comprises: collecting unit is used to gather Target Photo and background picture as the training pictures; Feature is asked for the unit, is used for concentrating the pixel that will have the predetermined image feature as unique point at the training picture, and asks for the distribution characteristics collection of unique point; And training unit, be used for that training obtains sorter to the distribution characteristics collection, wherein, the distribution characteristics collection comprises directivity entropy feature at least.
In addition, feature is asked for the unit and is used for Target Photo and background picture that the training picture is concentrated are divided into the zone of any size and shape, and obtains the distribution characteristics of each unique point in each zone as the distribution characteristics collection.
And the predetermined image feature comprises following at least a: statistical nature comprises average, the variance and covariance at brightness, gradient, high-order gradient, color and edge at least; Response comprises brightness, gradient, high-order gradient, color and the edge response to various wave filters at least; And the linear or nonlinear combination of the characteristics of image in zone.
The distribution characteristics of unique point in each zone comprises following at least a: at least a in brightness, gradient, high-order gradient, color, the distribution situation of edge in each zone; At least a in the distribution situation of the gradient of all directions, high-order gradient; And at least a in the joint distribution situation of each zone and all directions of gradient, high-order gradient.
In addition, feature is asked for the unit and comprised: unique point coordinate distribution characteristics is asked for the unit, the zone is divided into the plurality of sub zone according to coordinate, the statistics coordinate figure is positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of subregion or unique point, thereby with the ratio situation of all subregion distribution characteristics as the unique point in zone.
And, feature is asked for the unit and also comprised: unique point coordinate distribution entropy feature is asked for the unit, the zone is divided into the plurality of sub zone according to coordinate, the statistics coordinate figure is positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of subregion or unique point, and asks for entropy that coordinate distributes to obtain the entropy feature of coordinate distribution.
In addition, feature is asked for the unit and also comprised: the unique point coordinate distribution arrangement entropy feature with directivity is asked for the unit, coordinate axis is rotated several directions, unique point is divided into several portions according to the direction of feature, the consistent unique point of the direction of statistical nature point and change in coordinate axis direction, and calculate entropy feature that consistent characteristic point coordinates distributes as directivity entropy feature.
Wherein, zone and subregion are rectangle or polygon; Overlap between the zone and between the subregion.
And unique point is the pixel that the predetermined image feature satisfies size, direction, span with the unique point with directivity, and coordinate axis is with arbitrarily angled rotation.
In above-mentioned training system, training unit is used for asking for distribution characteristics that the unit seeks out from feature and concentrates and choose one or more characteristics of image of active zone partial objectives for picture and background picture as the validity feature collection; Characteristics of image or the characteristics of image concentrated by Boosting method training validity feature make up corresponding Weak Classifier, and distribute different weights for each Weak Classifier, thereby combination obtains final sorter, or by the SVM method validity feature collection is unified training to obtain sorter.
In addition, training system also comprises: training unit again is used for training again through the background picture that is mistaken as Target Photo after the training.
Wherein, the distribution characteristics collection at least also comprises HoG feature and/or Harr-like feature.
According to a further aspect in the invention, provide to be used at image or video by training the training method of the sorter that obtains to distinguish target and background, it comprises: acquisition step, gather Target Photo and background picture as the training pictures; Feature is asked for step, concentrates the pixel that will have the predetermined image feature as unique point at the training picture, and asks for the distribution characteristics collection of unique point; And training step, training obtains sorter to the distribution characteristics collection, and wherein, the distribution characteristics collection comprises directivity entropy feature at least.
In addition, ask in the step in feature, Target Photo and background picture that the training picture is concentrated are divided into the zone of any size and shape, and obtain the distribution characteristics of each unique point in each zone as the distribution characteristics collection.
And the predetermined image feature comprises following at least a: statistical nature comprises average, the variance and covariance at brightness, gradient, high-order gradient, color and edge at least; Response comprises brightness, gradient, high-order gradient, color and the edge response to various wave filters at least; And the linear or nonlinear combination of the characteristics of image in zone.
The distribution characteristics of unique point in each zone comprises following at least a: at least a in brightness, gradient, high-order gradient, color, the distribution situation of edge in each zone; At least a in the distribution situation of the gradient of all directions, high-order gradient; And at least a in the joint distribution situation of each zone and all directions of gradient, high-order gradient.
In addition, ask in the step in feature and to comprise: unique point coordinate distribution characteristics is asked for step, the zone is divided into the plurality of sub zone according to coordinate, the statistics coordinate figure is positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of subregion or unique point, thereby with the ratio situation of all subregion distribution characteristics as the unique point in zone.
And, ask in the step in feature and to comprise: unique point coordinate distribution entropy feature is asked for step, the zone is divided into the plurality of sub zone according to coordinate, the statistics coordinate figure is positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of subregion or unique point, and asks for entropy that coordinate distributes to obtain the entropy feature of coordinate distribution.
In addition, ask in the step in feature and to comprise: the unique point coordinate distribution arrangement entropy feature with directivity is asked for step, coordinate axis is rotated several directions, unique point is divided into several portions according to the direction of feature, the consistent unique point of the direction of statistical nature point and change in coordinate axis direction, and calculate entropy feature that consistent characteristic point coordinates distributes as directivity entropy feature.
Wherein, zone and subregion are rectangle or polygon; Overlap between the zone and between the subregion.
And unique point is the pixel that the predetermined image feature satisfies size, direction, span with the unique point with directivity, and coordinate axis is with arbitrarily angled rotation.
In above-mentioned training method, in above-mentioned training step, the distribution characteristics that seeks out the step is concentrated chooses one or more characteristics of image of active zone partial objectives for picture and background picture as the validity feature collection from asking in feature; Characteristics of image or the characteristics of image concentrated by Boosting method training validity feature make up corresponding Weak Classifier, and distribute different weights for each Weak Classifier, thereby combination obtains final sorter, or by the SVM method validity feature collection is unified training to obtain sorter.
In addition, training method also comprises: training step again, and to training again through the background picture that is mistaken as Target Photo after the training.
Wherein, the distribution characteristics collection at least also comprises HoG feature and/or Harr-like feature.
According to other aspects of the invention, provide user tropism's entropy feature at least in image or video, target image to be carried out the detection method of target detection, it comprises: step 1, in a two field picture of image that needs detect or video, whether the destination object that detects the certain size class in the optional position of image exists; Step 2, under the situation that destination object exists, the probability that the position of the target image of direction of passage entropy representative record size and the target of size exist in the position, thereby the probability distribution of destination object location size in the acquisition image; And step 3, according to summary distribution the carrying out position of aftertreatment of destination object location to judge whether destination object exists and exist.
Wherein, in step 2, may further comprise the steps: the image of position dimension is asked for directivity entropy feature at least, thereby obtain the feature set of the image of position dimension; And be probability from the feature set of target image acquisition by the classifier calculated feature set.
In addition, in step 2, image is asked for directivity entropy feature, HoG feature and/or Harr-like feature, thereby obtain the feature set of the image of position dimension.
In above-mentioned detection method, if content to be detected is a still image, then whether the target that detects in the certain size scope in the optional position of still image exists, when in scope, having target, if the position exists the target image of size greater than first threshold, the position dimension at record object place and ask for the probability that the position dimension target image exists then, thus the probability distribution of target location size in this image obtained; According to the probability distribution of target location carry out aftertreatment ask for final testing result, be target existence whether and the position in still image; If content to be detected is a video, then each frame video is considered as a width of cloth still image and detects.
Wherein, the acquiring method of probability comprises: the various characteristics of image that adopt when the image of position dimension is asked for training obtain the feature set of the image of position dimension; Adopting sorter to come the calculated characteristics collection is that just the image of position dimension is the probability of target image from the probability of the feature set of target image acquisition.
In addition, when judge detecting when representing same target greater than the image of the close positions of first threshold and size, in aftertreatment, merging close position and size and obtaining net result for the probability of target image.
And, the method that merges comprises: the weighted mean value according to probability of getting close positions and size is asked for position and the size after the merging, probability after the merging is the weighted mean value of the probability of close positions and size, the probability after the merging greater than the amalgamation result of second threshold value as net result; On probability distribution graph, ask for the maximum value of probability, maximum value greater than the position of the 3rd threshold value as net result.
When above-mentioned content to be detected is video, on the basis that above-mentioned still image detects, carry out joint-detection with reference to the correlativity of the above-mentioned video of each frame.
Therefore, adopt the present invention, can be in image or video data recognition target image existence whether with the position that exists, and can improve correct rate of target detection.
Other features and advantages of the present invention will be set forth in the following description, and, partly from instructions, become apparent, perhaps understand by implementing the present invention.Purpose of the present invention and other advantages can realize and obtain by specifically noted structure in the instructions of being write, claims and accompanying drawing.
Description of drawings
Accompanying drawing is used to provide further understanding of the present invention, and constitutes the part of instructions, is used from explanation the present invention with embodiments of the invention one, is not construed as limiting the invention.In the accompanying drawings:
Fig. 1 is the block diagram according to the structure of training system of the present invention;
Fig. 2 is the process flow diagram according to training method of the present invention;
Fig. 3 is according to the Target Photo of the embodiment of the invention and background picture and the synoptic diagram that picture is divided into zone and subregion;
Fig. 4 is the process flow diagram that carries out the training method of target detection according to the employing feature of the embodiment of the invention in video or image;
Fig. 5 is the synoptic diagram according to the HoG feature of the embodiment of the invention;
Fig. 6 is according to the directivity entropy characteristic principle of the embodiment of the invention and the synoptic diagram of computing method, Fig. 6 (a) is former picture figure, Fig. 6 (b) is a gradient map, Fig. 6 (c) is near the gradient component of gradient direction 0 degree, Fig. 6 (d) is near the gradient component of gradient direction 45 degree, Fig. 6 (e) is near the gradient component of gradient direction 90 degree, Fig. 6 (f) is near the gradient component of gradient direction 135 degree, Fig. 6 (g) is the figure of rotatable coordinate axis, Fig. 6 (h) is the coordinate diagram of feature points, and Fig. 6 (i) is to calculate the synoptic diagram of entropy to new coordinate axis projection;
Fig. 7 is the process flow diagram according to detection method of the present invention; And
Fig. 8 is the process flow diagram that carries out the detection method of target detection according to the employing feature of the embodiment of the invention in video or image.
Embodiment
Below in conjunction with accompanying drawing the preferred embodiments of the present invention are described, should be appreciated that preferred embodiment described herein only is used for description and interpretation the present invention, and be not used in qualification the present invention.
Fig. 1 is the block diagram according to the structure of training system 100 of the present invention.
As shown in Figure 1, be used for comprising by the training system 100 of training the sorter that obtains to distinguish target and background at image or video: collecting unit 102 is used to gather Target Photo and background picture as the training pictures; Feature is asked for unit 104, is used for concentrating the pixel that will have the predetermined image feature as unique point at the training picture, and asks for the distribution characteristics collection of unique point; And training unit 106, be used for that training obtains sorter to the distribution characteristics collection, wherein, the distribution characteristics collection comprises directivity entropy feature at least.
In addition, feature is asked for unit 104 and is used for Target Photo and background picture that the training picture is concentrated are divided into the zone of any size and shape, and obtains the distribution characteristics of each unique point in each zone as the distribution characteristics collection.
And the predetermined image feature comprises following at least a: statistical nature comprises average, the variance and covariance at brightness, gradient, high-order gradient, color and edge at least; Response comprises brightness, gradient, high-order gradient, color and the edge response to various wave filters at least; And the linear or nonlinear combination of the characteristics of image in zone.
The distribution characteristics of unique point in each zone comprises following at least a: at least a in brightness, gradient, high-order gradient, color, the distribution situation of edge in each zone; At least a in the distribution situation of the gradient of all directions, high-order gradient; And at least a in the joint distribution situation of each zone and all directions of gradient, high-order gradient.
In addition, feature is asked for unit 104 and comprised: unique point coordinate distribution characteristics is asked for the unit, the zone is divided into the plurality of sub zone according to coordinate, the statistics coordinate figure is positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of subregion or unique point, thereby with the ratio situation of all subregion distribution characteristics as the unique point in zone.
And, feature is asked for unit 104 and also comprised: unique point coordinate distribution entropy feature is asked for the unit, the zone is divided into the plurality of sub zone according to coordinate, the statistics coordinate figure is positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of subregion or unique point, and asks for entropy that coordinate distributes to obtain the entropy feature of coordinate distribution.
In addition, feature is asked for unit 104 and also comprised: the unique point coordinate distribution arrangement entropy feature with directivity is asked for the unit, coordinate axis is rotated several directions, unique point is divided into several portions according to the direction of feature, the consistent unique point of the direction of statistical nature point and change in coordinate axis direction, and calculate entropy feature that consistent characteristic point coordinates distributes as directivity entropy feature.
Wherein, zone and subregion are rectangle or polygon; Overlap between the zone and between the subregion.
And unique point is the pixel that the predetermined image feature satisfies size, direction, span with the unique point with directivity, and coordinate axis is with arbitrarily angled rotation.
In above-mentioned training system 100, training unit 106 is used for asking for distribution characteristics that unit 104 seeks out from feature and concentrates and choose one or more characteristics of image of active zone partial objectives for picture and background picture as the validity feature collection; Characteristics of image or the characteristics of image concentrated by Boosting method training validity feature make up corresponding Weak Classifier, and distribute different weights for each Weak Classifier, thereby combination obtains final sorter, or by the SVM method validity feature collection is unified training to obtain sorter.Because target has various changes, thus the sorter that is obtained generally be several, promptly one or more.The output of training system is sorter.
In addition, training system 100 also comprises: training unit again is used for training again through the background picture that is mistaken as Target Photo after the training.This process of asking for again, training generally will experience twice or thrice.
Wherein, the distribution characteristics collection at least also comprises HoG feature and/or Harr-like feature.
Fig. 2 is the process flow diagram according to training method of the present invention.
As shown in Figure 2, be used for comprising by the training method of training the sorter that obtains to distinguish target and background at image or video: acquisition step S202, gather Target Photo and background picture as the training pictures; Feature is asked for step S204, concentrates the pixel that will have the predetermined image feature as unique point at the training picture, and asks for the distribution characteristics collection of unique point; And training step S206, training obtains sorter to the distribution characteristics collection, and wherein, the distribution characteristics collection comprises directivity entropy feature at least.
In addition, ask among the step S204 in feature, Target Photo and background picture that the training picture is concentrated are divided into the zone of any size and shape, and obtain the distribution characteristics of each unique point in each zone as the distribution characteristics collection.
And the predetermined image feature comprises following at least a: statistical nature comprises average, the variance and covariance at brightness, gradient, high-order gradient, color and edge at least; Response comprises brightness, gradient, high-order gradient, color and the edge response to various wave filters at least; And the linear or nonlinear combination of the characteristics of image in zone.
The distribution characteristics of unique point in each zone comprises following at least a: at least a in brightness, gradient, high-order gradient, color, the distribution situation of edge in each zone; At least a in the distribution situation of the gradient of all directions, high-order gradient; And at least a in the joint distribution situation of each zone and all directions of gradient, high-order gradient.
In addition, ask among the step S204 in feature and to comprise: unique point coordinate distribution characteristics is asked for step, the zone is divided into the plurality of sub zone according to coordinate, the statistics coordinate figure is positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of subregion or unique point, thereby with the ratio situation of all subregion distribution characteristics as the unique point in zone.
And, ask among the step S204 in feature and to comprise: unique point coordinate distribution entropy feature is asked for step, the zone is divided into the plurality of sub zone according to coordinate, the statistics coordinate figure is positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of subregion or unique point, and asks for entropy that coordinate distributes to obtain the entropy feature of coordinate distribution.
In addition, ask among the step S204 in feature and to comprise: the unique point coordinate distribution arrangement entropy feature with directivity is asked for step, coordinate axis is rotated several directions, unique point is divided into several portions according to the direction of feature, the consistent unique point of the direction of statistical nature point and change in coordinate axis direction, and calculate entropy feature that consistent characteristic point coordinates distributes as directivity entropy feature.
Wherein, zone and subregion are rectangle or polygon; Overlap between the zone and between the subregion.
And unique point is the pixel that the predetermined image feature satisfies size, direction, span with the unique point with directivity, and coordinate axis is with arbitrarily angled rotation.
In above-mentioned training method, in above-mentioned training step S206, the distribution characteristics that seeks out the step S204 is concentrated chooses one or more characteristics of image of active zone partial objectives for picture and background picture as the validity feature collection from asking in feature; Characteristics of image or the characteristics of image concentrated by Boosting method training validity feature make up corresponding Weak Classifier, and distribute different weights for each Weak Classifier, thereby combination obtains final sorter, or by the SVM method validity feature collection is unified training to obtain sorter.Because target has various changes, thus the sorter that is obtained generally be several, promptly one or more.The output of training system is sorter.
In addition, training method also comprises: training step again, and to training again through the background picture that is mistaken as Target Photo after the training.This process of asking for again, training generally will experience twice or thrice.
Wherein, the distribution characteristics collection at least also comprises HoG feature and/or Harr-like feature.
Particularly, Fig. 3 is according to the Target Photo of the embodiment of the invention and background picture and the synoptic diagram that picture is divided into zone and subregion.In Fig. 3, be example with the human body picture as the situation of Target Photo, but be not limited in this.Wherein, the picture in the training set is divided into the zone of any size and shape, the shape in zone can be overlapped between the zone mutually for rectangle or other polygons.
Fig. 4 is the process flow diagram that carries out the training method of target detection according to the employing feature of the embodiment of the invention in video or image.
As shown in Figure 4, training method may further comprise the steps:
Step S402 gathers Target Photo and background picture as the training pictures;
Step S404, Target Photo and background picture that the training picture is concentrated are asked for various features, the zone that Target Photo that the training picture is concentrated and background picture are divided into any size and arbitrary shape, should be divided into the plurality of sub zone according to coordinate in the zone, the statistics coordinate figure is positioned at the distribution characteristics of the ratio situation of the ratio of characteristic quantity summation of number ratio, unique point of the unique point of subregion or all subregion as this provincial characteristics point;
Step S406 instructs the distribution characteristics of asking for and to obtain sorter; And
Step S408, training step asks for, trains the background picture that is mistaken as Target Photo to ask for again, train to process again.
Particularly, be example: will train the Target Photo of pictures to normalize to identical size, and ask for the directivity entropy feature of HoG feature and gradient with following situation.
The brightness of setting image be I (x, y),
Setting the horizontal direction gradient is I
x(x, y)=d (I (x, y))/dx=I (x+1, y)-I (x-1, y),
The setting vertical gradient is I
y(x, y)=d (I (x, y))/dy=I (x, y+1)-I (x, y-1),
The setting gradient is
The setting gradient direction be θ (x, y)=argtg (| I
y/ I
x|),
In addition, three color components that also can each pixel calculate respectively x, the differential value of y, and its maximum value is as the gradient of this pixel.
And, as shown in Figure 5, picture is divided into some, be divided into four fritters with every then, gradient map is divided into several directions, ask for the summation of the gradient of each direction of all pixels in each fritter, then the gradient summation of all directions of fritter in every is carried out normalization, obtain the HoG feature of this piece.
In addition, Fig. 6 is according to the directivity entropy characteristic principle of the embodiment of the invention and the synoptic diagram of computing method, Fig. 6 (a) is former picture figure, Fig. 6 (b) is a gradient map, Fig. 6 (c) is near the gradient component of gradient direction 0 degree, Fig. 6 (d) is near the gradient component of gradient direction 45 degree, Fig. 6 (e) is near the gradient component of gradient direction 90 degree, Fig. 6 (f) is near the gradient component of gradient direction 135 degree, Fig. 6 (g) is the figure of rotatable coordinate axis, Fig. 6 (h) is the coordinate diagram of feature points, and Fig. 6 (i) is to calculate the synoptic diagram of entropy to new coordinate axis projection.The directivity entropy feature of gradient as shown in Figure 6.Gradient map is divided into several directions, in the gradient component, appoints and get a piece, calculate the directivity entropy feature of the gradient of this this piece of direction.At first the coordinate axis with the gradient component is rotated according to this component direction.(x, some y) are (x ', y ') at the coordinate of new coordinate axis in former figure coordinate.
As shown in Figure 6, the angle of new coordinate axis and former coordinate axis is a, and (x is s=sqrt (x^2+y^2) from the initial point distance y) to point, and the angle that line and former coordinate axis form between this point and initial point is b=argtg (y/x), then this new coordinate y '=s*sin (b-a).
For starting point is (x
0, y
0), width is the w pixel, highly is the directivity gradient piece of h pixel, the pixel that all gradient directions are consistent with this gradient component direction is to y ' axial projection.The distribution of directivity gradient in piece can be represented by the distribution proportion on new coordinate axis of this direction gradient weighting of all pixels in the piece:
In addition, the directivity entropy of the gradient of a direction then is in this piece:
In the various features of extracting, choose can distinguish target and background certain characteristics as the validity feature collection.That is, marginal point in this piece to new coordinate axis projection, is calculated the distribution at y ' axle, thereby obtains the entropy of Gradient distribution in this direction.
The training validity feature is concentrated the Weak Classifier of feature or characteristics combination correspondence, adopts the method for boosting to select Weak Classifier, and each Weak Classifier weight of reasonable distribution is combined into final sorter.
Fig. 7 is the process flow diagram according to detection method of the present invention.
As shown in Figure 7, at least user tropism's entropy feature comprises the detection method that target image carries out target detection in image or video: step S702, in a two field picture of image that needs detect or video, whether the destination object that detects the certain size class in the optional position of image exists; Step S704, under the situation that destination object exists, the probability that the position of the target image of direction of passage entropy representative record size and the target of size exist in the position, thereby the probability distribution of destination object location size in the acquisition image; And step S706, according to summary distribution the carrying out position of aftertreatment of destination object location to judge whether destination object exists and exist.
Wherein, in step S704, may further comprise the steps: the image of position dimension is asked for directivity entropy feature at least, thereby obtain the feature set of the image of position dimension; And be probability from the feature set of target image acquisition by the classifier calculated feature set.
In addition, in step S704, image is asked for directivity entropy feature, HoG feature and/or Harr-like feature, thereby obtain the feature set of the image of position dimension.
In detection method, if content to be detected is a still image, then whether the target that detects in the certain size scope in the optional position of still image exists, when in scope, having target, if the position exists the target image of size greater than first threshold, the position dimension at record object place and ask for the probability that the position dimension target image exists then, thus the probability distribution of target location size in this image obtained; According to the probability distribution of target location carry out aftertreatment ask for final testing result, be target existence whether and the position in still image; If content to be detected is a video, then each frame video is considered as a width of cloth still image and detects.
Wherein, the acquiring method of probability comprises: the various characteristics of image that adopt when the image of position dimension is asked for training obtain the feature set of the image of position dimension; Adopting sorter to come the calculated characteristics collection is that just the image of position dimension is the probability of target image from the probability of the feature set of target image acquisition.
In addition, when judge detecting when representing same target greater than the image of the close positions of first threshold and size, in aftertreatment, merging close position and size and obtaining net result for the probability of target image.
And, the method that merges comprises: the weighted mean value according to probability of getting close positions and size is asked for position and the size after the merging, probability after the merging is the weighted mean value of the probability of close positions and size, the probability after the merging greater than the amalgamation result of second threshold value as net result; On probability distribution graph, ask for the maximum value of probability, maximum value greater than the position of the 3rd threshold value as net result.
When content to be detected is video, on the basis that still image detects, carry out joint-detection with reference to the correlativity of each frame video.
Particularly, Fig. 8 is the process flow diagram that carries out the detection method of target detection according to the employing feature of the embodiment of the invention in video or image.
As shown in Figure 8, be input as a certain two field picture 501 in image to be detected or the video.May further comprise the steps:
Step S802 imports image to be detected or a certain two field picture in the video, the image in the acquisition image in the search window of optional position arbitrary dimension.
Step S804, when changing the size of search window, size of images changes thereupon in the window, a perhaps mobile search position of window, the window size size that maintains the standard, and the image that original image changes behind the various different sizes is also searched for.Picture size in the search window that obtains is like this fixed, but content sources is in the interpolation of original image element.
Step S806, calculate image in the search window with training the time the same various features calculated.If the size of search window changes,, calculate after the picture size in the window can being zoomed to standard size for the identical feature of the image calculation of different size.Perhaps the size that directly changes feature is calculated.Search for if original image is changed various different sizes, the image in the search window is a standard size, can directly carry out feature calculation.
Step S808 confirms that by sorter each search window has the probability of the existence of target.Obtaining all search window images is the Probability p of target
i(i=0......n, n are total number of search window).With p
iEach position of window size of＞threshold_1 (xi, yi, ri) and this probability note.
Step S810 is by the p that notes
iEach position of window size of＞T1 is carried out aftertreatment and is obtained the final location of target in former figure.
Detection probability may be represented same target greater than the close positions of first threshold T1 and the image of size, so aftertreatment can merge close position and size obtains net result.
A kind of simple merging method is with first the window's position size (x0 in the record, y0, r0) be used as initial results (x_mode0, y_mode0, r_mode0), seeking next window position dimension and its difference exists | xi-x_mode0|＜difx*r0, | yi-y_mode0|＜difh*r0, | log (ri)-log (r_mode0) | the record the within＜difr scope.Merge the renewal result by these two records according to detection probability, that is:
Merge all windows close according to the method described above and obtain first result with first the window's position size.Merge according to similar method in remaining record the inside, obtain all possible result.
P by Fused window
iThe position that obtains after this merging of combination be the probability P of target
j, P
jThink in the time of＞T2 that this position is a target.
The output result is all P
jThe position of＞T2 (T2 is second threshold value), i.e. detected target location in former figure.
Be the preferred embodiments of the present invention only below, be not limited to the present invention, for a person skilled in the art, the present invention can have various changes and variation.Within the spirit and principles in the present invention all, any modification of being done, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.
Claims (27)
1. training system is used at image or video by training the sorter that obtains to distinguish target and background, and described training system is characterised in that, comprising:
Collecting unit is used to gather Target Photo and background picture as the training pictures;
Feature is asked for the unit, is used for concentrating the pixel that will have the predetermined image feature as unique point at described training picture, and asks for the distribution characteristics collection of described unique point; And
Training unit is used for that training obtains described sorter to described distribution characteristics collection,
Wherein, described distribution characteristics collection comprises directivity entropy feature at least.
2. training system according to claim 1 is characterized in that,
Described feature is asked for the unit and is used for described Target Photo and described background picture that described training picture is concentrated are divided into the zone of any size and shape, and obtains the distribution characteristics of each described unique point in each described zone as described distribution characteristics collection.
3. training system according to claim 2 is characterized in that, described predetermined image feature comprises following at least a:
Statistical nature comprises average, the variance and covariance at brightness, gradient, high-order gradient, color and edge at least;
Response comprises brightness, gradient, high-order gradient, color and the edge response to various wave filters at least; And
Linear or the nonlinear combination of the characteristics of image in described zone.
4. training system according to claim 2 is characterized in that, the distribution characteristics of described unique point in described each zone comprises following at least a:
At least a in brightness, gradient, high-order gradient, color, the distribution situation of edge in described each zone;
At least a in the distribution situation of the gradient of all directions, high-order gradient; And
At least a in the joint distribution situation of each zone and all directions of gradient, high-order gradient.
5. training system according to claim 2 is characterized in that,
Described feature is asked for the unit and comprised: unique point coordinate distribution characteristics is asked for the unit, described zone is divided into the plurality of sub zone according to coordinate, add up described coordinate figure and be positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of described subregion or described unique point, thereby with the ratio situation of the described all subregion distribution characteristics as the unique point in described zone.
6. training system according to claim 5 is characterized in that,
Described feature is asked for the unit and also comprised: unique point coordinate distribution entropy feature is asked for the unit, described zone is divided into described plurality of sub zone according to coordinate, add up described coordinate figure and be positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of described subregion or described unique point, and ask for entropy that coordinate distributes to obtain the entropy feature that described coordinate distributes.
7. training system according to claim 6 is characterized in that,
Described feature is asked for the unit and also comprised: the unique point coordinate distribution arrangement entropy feature with directivity is asked for the unit, described coordinate axis is rotated several directions, described unique point is divided into several portions according to the direction of feature, add up the direction and the consistent unique point of described change in coordinate axis direction of described unique point, and calculate entropy feature that described consistent characteristic point coordinates distributes as directivity entropy feature.
8. according to each described training system in the claim 1 to 7, it is characterized in that,
Described zone and described subregion are rectangle or polygon;
Overlap between the described zone and between the described subregion.
9. according to each described training system in the claim 5 to 7, it is characterized in that,
Described unique point and described unique point with directivity are the pixels that described predetermined image feature satisfies size, direction, span, and
Described coordinate axis is with arbitrarily angled rotation.
10. training system according to claim 1 is characterized in that,
Described training unit is used for asking for described distribution characteristics that the unit seeks out from described feature and concentrates and choose active zone and divide one or more characteristics of image of described Target Photo and described background picture as the validity feature collection;
The described characteristics of image or the characteristics of image of training described validity feature to concentrate by the Boosting method make up corresponding Weak Classifier, and distribute different weights for each Weak Classifier, thereby combination obtains final sorter, or by the SVM method described validity feature collection is unified training to obtain sorter.
11. training system according to claim 1 is characterized in that, described training system also comprises:
Again training unit is used for training again through the described background picture that is mistaken as described Target Photo after the training.
12. training system according to claim 1 is characterized in that,
Described distribution characteristics collection at least also comprises HoG feature and/or Harr-like feature.
13. a training method is used at image or video by training the sorter that obtains to distinguish target and background, described training method is characterised in that, comprising:
Acquisition step is gathered Target Photo and background picture as the training pictures;
Feature is asked for step, concentrates the pixel that will have the predetermined image feature as unique point at described training picture, and asks for the distribution characteristics collection of described unique point; And
Training step, training obtains described sorter to described distribution characteristics collection,
Wherein, described distribution characteristics collection comprises directivity entropy feature at least.
14. training method according to claim 13 is characterized in that,
Ask in the step in described feature, described Target Photo and described background picture that described training picture is concentrated are divided into the zone of any size and shape, and obtain the distribution characteristics of each described unique point in each described zone as described distribution characteristics collection.
15. training method according to claim 14 is characterized in that, described predetermined image feature comprises following at least a:
Statistical nature comprises average, the variance and covariance at brightness, gradient, high-order gradient, color and edge at least;
Response comprises brightness, gradient, high-order gradient, color and the edge response to various wave filters at least; And
Linear or the nonlinear combination of the characteristics of image in described zone.
16. training method according to claim 14 is characterized in that, the distribution characteristics of described unique point in described each zone comprises following at least a:
At least a in brightness, gradient, high-order gradient, color, the distribution situation of edge in described each zone;
At least a in the distribution situation of the gradient of all directions, high-order gradient; And
At least a in the joint distribution situation of each zone and all directions of gradient, high-order gradient.
17. training method according to claim 14 is characterized in that,
Ask in the step in described feature and to comprise: unique point coordinate distribution characteristics is asked for step, described zone is divided into the plurality of sub zone according to coordinate, add up described coordinate figure and be positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of described subregion or described unique point, thereby with the ratio situation of the described all subregion distribution characteristics as the unique point in described zone.
18. training method according to claim 17 is characterized in that,
Ask in the step in described feature and to comprise: unique point coordinate distribution entropy feature is asked for step, described zone is divided into described plurality of sub zone according to coordinate, add up described coordinate figure and be positioned at the ratio of the characteristic quantity summation of the number ratio of unique point of described subregion or described unique point, and ask for entropy that coordinate distributes to obtain the entropy feature that described coordinate distributes.
19. training method according to claim 18 is characterized in that,
Ask in the step in described feature and to comprise: the unique point coordinate distribution arrangement entropy feature with directivity is asked for step, described coordinate axis is rotated several directions, described unique point is divided into several portions according to the direction of feature, add up the direction and the consistent unique point of described change in coordinate axis direction of described unique point, and calculate entropy feature that described consistent characteristic point coordinates distributes as directivity entropy feature.
20. according to each described training method in the claim 13 to 19, it is characterized in that,
Described zone and described subregion are rectangle or polygon;
Overlap between the described zone and between the described subregion.
21. according to each described training method in the claim 17 to 19, it is characterized in that,
Described unique point and described unique point with directivity are the pixels that described predetermined image feature satisfies size, direction, span, and
Described coordinate axis is with arbitrarily angled rotation.
22. training method according to claim 13 is characterized in that,
In described training step, concentrate and choose active zone and divide one or more characteristics of image of described Target Photo and described background picture as the validity feature collection from ask for the described distribution characteristics that seeks out the step in described feature;
The described characteristics of image or the characteristics of image of training described validity feature to concentrate by the Boosting method make up corresponding Weak Classifier, and distribute different weights for each Weak Classifier, thereby combination obtains final sorter, or by the SVM method described validity feature collection is unified training to obtain sorter.
23. training method according to claim 13 is characterized in that, described training method also comprises:
Again training step is trained again to the described background picture that is mistaken as described Target Photo after the process training.
24. training method according to claim 13 is characterized in that,
Described distribution characteristics collection at least also comprises HoG feature and/or Harr-like feature.
25. a detection method, user tropism's entropy feature is carried out target detection to target image in image or video at least, it is characterized in that, described detection method comprises:
Step 1, in a two field picture of described image that needs detect or described video, whether the destination object that detects the certain size class in the optional position of described image exists;
Step 2, under the situation that described destination object exists, the probability that the position of the described target image by the described size of described directivity entropy representative record and the described target of described size exist in described position, thus the probability distribution of the size of destination object location described in the described image obtained; And
Step 3 is according to summary distribution the carrying out position of aftertreatment to judge whether described destination object exists and exist of described destination object location.
26. detection method according to claim 25 is characterized in that,
In described step 2, may further comprise the steps:
The image of described position dimension is asked for described directivity entropy feature at least, thereby obtain the feature set of the described image of described position dimension; And
It by the described feature set of classifier calculated the probability of the feature set that obtains from described target image.
27. detection method according to claim 25 is characterized in that,
In described step 2, described image is asked for described directivity entropy feature, HoG feature and/or Harr-like feature, thereby obtain the feature set of the image of described position dimension.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN200810210129A CN101655910A (en) | 2008-08-21 | 2008-08-21 | Training system, training method and detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN200810210129A CN101655910A (en) | 2008-08-21 | 2008-08-21 | Training system, training method and detection method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN101655910A true CN101655910A (en) | 2010-02-24 |
Family
ID=41710196
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN200810210129A Pending CN101655910A (en) | 2008-08-21 | 2008-08-21 | Training system, training method and detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN101655910A (en) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101814146A (en) * | 2010-04-20 | 2010-08-25 | 杭州电子科技大学 | Method for extracting action potential feature based on multiple wavelet transformation |
| CN101976345A (en) * | 2010-09-30 | 2011-02-16 | 哈尔滨工程大学 | Method for recognizing image scale invariant pattern under noise condition |
| CN102184388A (en) * | 2011-05-16 | 2011-09-14 | 苏州两江科技有限公司 | Face and vehicle adaptive rapid detection system and detection method |
| CN102930287A (en) * | 2012-09-26 | 2013-02-13 | 上海理工大学 | Overlook-based detection and counting system and method for pedestrians |
| CN103324937A (en) * | 2012-03-21 | 2013-09-25 | 日电（中国）有限公司 | Method and device for labeling targets |
| CN104268595A (en) * | 2014-09-24 | 2015-01-07 | 深圳市华尊科技有限公司 | General object detecting method and system |
| CN105496563A (en) * | 2015-12-04 | 2016-04-20 | 上海联影医疗科技有限公司 | Method for calibrating medical image positioning line |
| CN105808610A (en) * | 2014-12-31 | 2016-07-27 | 中国科学院深圳先进技术研究院 | Internet picture filtering method and device |
| CN103345631B (en) * | 2013-06-04 | 2016-12-28 | 北京大学深圳研究生院 | Image characteristics extraction, training, detection method and module, device, system |
| CN108010012A (en) * | 2017-10-26 | 2018-05-08 | 厦门市智联信通物联网科技有限公司 | Tunnel portal black-hole effect evaluation method, system and intelligent lighting system |
| CN111753870A (en) * | 2020-04-16 | 2020-10-09 | 杭州海康威视数字技术股份有限公司 | Training method and device of target detection model and storage medium |
| CN112184663A (en) * | 2020-09-27 | 2021-01-05 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting foreign matter of anti-snaking shock absorber mounting seat of railway motor car |
| CN112639396A (en) * | 2019-08-07 | 2021-04-09 | 株式会社日立高新技术 | Dimension measuring apparatus, dimension measuring method, and semiconductor manufacturing system |
-
2008
- 2008-08-21 CN CN200810210129A patent/CN101655910A/en active Pending
Non-Patent Citations (2)
| Title |
|---|
| LI HAI-YAN等: "Face detection with Bayesian classifier", 《云南大学学报（自然科学版）》 * |
| 张新征等: "SAR目标多尺度概率密度估计与识别", 《系统工程与电子技术》 * |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101814146B (en) * | 2010-04-20 | 2013-01-09 | 杭州电子科技大学 | Method for extracting action potential feature based on multiple wavelet transformation |
| CN101814146A (en) * | 2010-04-20 | 2010-08-25 | 杭州电子科技大学 | Method for extracting action potential feature based on multiple wavelet transformation |
| CN101976345A (en) * | 2010-09-30 | 2011-02-16 | 哈尔滨工程大学 | Method for recognizing image scale invariant pattern under noise condition |
| CN101976345B (en) * | 2010-09-30 | 2013-02-06 | 哈尔滨工程大学 | Method for recognizing image scale invariant pattern under noise condition |
| CN102184388A (en) * | 2011-05-16 | 2011-09-14 | 苏州两江科技有限公司 | Face and vehicle adaptive rapid detection system and detection method |
| CN103324937A (en) * | 2012-03-21 | 2013-09-25 | 日电（中国）有限公司 | Method and device for labeling targets |
| CN103324937B (en) * | 2012-03-21 | 2016-08-03 | 日电（中国）有限公司 | The method and apparatus of label target |
| CN102930287A (en) * | 2012-09-26 | 2013-02-13 | 上海理工大学 | Overlook-based detection and counting system and method for pedestrians |
| CN102930287B (en) * | 2012-09-26 | 2015-09-02 | 上海理工大学 | A kind of detection number system and method for overlooking pedestrian |
| CN103345631B (en) * | 2013-06-04 | 2016-12-28 | 北京大学深圳研究生院 | Image characteristics extraction, training, detection method and module, device, system |
| CN104268595A (en) * | 2014-09-24 | 2015-01-07 | 深圳市华尊科技有限公司 | General object detecting method and system |
| CN104268595B (en) * | 2014-09-24 | 2018-02-13 | 深圳市华尊科技股份有限公司 | general object detection method and system |
| CN105808610A (en) * | 2014-12-31 | 2016-07-27 | 中国科学院深圳先进技术研究院 | Internet picture filtering method and device |
| CN105496563A (en) * | 2015-12-04 | 2016-04-20 | 上海联影医疗科技有限公司 | Method for calibrating medical image positioning line |
| CN105496563B (en) * | 2015-12-04 | 2018-05-04 | 上海联影医疗科技有限公司 | The method for demarcating backbone medical image position line |
| CN108010012A (en) * | 2017-10-26 | 2018-05-08 | 厦门市智联信通物联网科技有限公司 | Tunnel portal black-hole effect evaluation method, system and intelligent lighting system |
| CN112639396A (en) * | 2019-08-07 | 2021-04-09 | 株式会社日立高新技术 | Dimension measuring apparatus, dimension measuring method, and semiconductor manufacturing system |
| CN111753870A (en) * | 2020-04-16 | 2020-10-09 | 杭州海康威视数字技术股份有限公司 | Training method and device of target detection model and storage medium |
| CN111753870B (en) * | 2020-04-16 | 2023-08-18 | 杭州海康威视数字技术股份有限公司 | Training method, device and storage medium of target detection model |
| CN112184663A (en) * | 2020-09-27 | 2021-01-05 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting foreign matter of anti-snaking shock absorber mounting seat of railway motor car |
| CN112184663B (en) * | 2020-09-27 | 2021-07-20 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting foreign matter of anti-snaking shock absorber mounting seat of railway motor car |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101655910A (en) | Training system, training method and detection method | |
| CN109977782B (en) | Cross-store operation behavior detection method based on target position information reasoning | |
| CN103824070B (en) | A kind of rapid pedestrian detection method based on computer vision | |
| CN101551852B (en) | Training system, training method and detection method | |
| CN102084397B (en) | Image processing device and method | |
| CN103700087B (en) | mobile detection method and device | |
| CN102947863B (en) | Moving object detection device | |
| US10789495B2 (en) | System and method for 1D root association providing sparsity guarantee in image data | |
| CN103810717B (en) | A kind of human body behavioral value method and device | |
| KR101476799B1 (en) | System and method for detecting object using depth information | |
| CN105046206B (en) | Based on the pedestrian detection method and device for moving prior information in video | |
| Zhang et al. | A pedestrian detection method based on SVM classifier and optimized Histograms of Oriented Gradients feature | |
| CN101872477A (en) | Method and device for detecting object in image and system containing device | |
| CN110210474A (en) | Object detection method and device, equipment and storage medium | |
| Bansod et al. | Transfer learning for video anomaly detection | |
| CN105404894A (en) | Target tracking method used for unmanned aerial vehicle and device thereof | |
| CN102521842B (en) | Method and device for detecting fast movement | |
| CN105354856A (en) | Human matching and positioning method and system based on MSER and ORB | |
| CN102479329A (en) | Classifier generation apparatus and method thereof, apparatus for detecting object in image and method thereof | |
| CN107403451A (en) | Adaptive binary feature monocular vision odometer method and computer, robot | |
| CN114926747A (en) | Remote sensing image directional target detection method based on multi-feature aggregation and interaction | |
| CN108010065A (en) | Low target quick determination method and device, storage medium and electric terminal | |
| Nizar et al. | Multi-object tracking and detection system based on feature detection of the intelligent transportation system | |
| CN104268595A (en) | General object detecting method and system | |
| CN105574889A (en) | Individual abnormal behavior detecting method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20100224 |