CN105354579A - Feature detection method and apparatus - Google Patents

Feature detection method and apparatus Download PDF

Info

Publication number
CN105354579A
CN105354579A CN201510727485.9A CN201510727485A CN105354579A CN 105354579 A CN105354579 A CN 105354579A CN 201510727485 A CN201510727485 A CN 201510727485A CN 105354579 A CN105354579 A CN 105354579A
Authority
CN
China
Prior art keywords
video image
eigenwert
gmm
sampled result
gaussian kernel
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.)
Granted
Application number
CN201510727485.9A
Other languages
Chinese (zh)
Other versions
CN105354579B (en
Inventor
毛敏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang Uniview Technologies Co Ltd
Original Assignee
Zhejiang Uniview Technologies Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang Uniview Technologies Co Ltd filed Critical Zhejiang Uniview Technologies Co Ltd
Priority to CN201510727485.9A priority Critical patent/CN105354579B/en
Publication of CN105354579A publication Critical patent/CN105354579A/en
Application granted granted Critical
Publication of CN105354579B publication Critical patent/CN105354579B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

The present invention provides a feature detection method and apparatus and aims at each to-be-detected video image. The method comprises: carrying out feature extraction on video images, and obtaining M eigenvalues; carrying out sampling on the M eigenvalues for N times, and obtaining N sampling results, wherein each sampling result comprises a part of eigenvalues in the M eigenvalues; aiming at each sampling result, adopting k Gaussian kernels for the sampling results, generating a GMM sub-model corresponding to the each sampling result, and obtaining N GMM sub-models; and after ranking the N GMM sub-models, obtaining corresponding GMM models, obtaining eigenvectors corresponding to the GMM models, and carrying out feature detection by using the eigenvectors. According to the technical scheme provided by the present invention, the corresponding GMM models can be generated by using a small number of K Gaussian kernels, thereby reducing calculation complexity of a feature detection algorithm, improving calculation performance, improving algorithm convergence, and effectively accelerating eigenvector generation.

Description

A kind of method and apparatus of feature detection
Technical field
The present invention relates to technical field of image processing, particularly relate to a kind of method and apparatus of feature detection.
Background technology
Feature detection has a wide range of applications field, such as video analysis, target detection, image recognition etc.Under normal circumstances, the process of feature detection comprises: a video image is divided into several subregions, feature extraction is carried out to every sub regions, obtain the eigenwert of every sub regions, finally the eigenwert of extraction is carried out cascade, generate one group of high dimensional feature vector, and by this high dimensional feature vector, video image is represented, thus based on the high dimensional feature vector of this video image, feature detection is carried out to this video image.
Because aforesaid way directly generates corresponding high dimensional feature vector based on original video image, it is in rotation, translation and when having larger noise, and reliability is lower.On the other hand, owing to directly adopting cascade system to generate high dimensional feature vector, feature inconsistent phenomenon can be produced when carrying out feature detection.
Summary of the invention
The invention provides a kind of method of feature detection, for each video image to be detected, said method comprising the steps of:
Feature extraction is carried out to video image, obtains M eigenwert; N sampling is carried out to M eigenwert, obtains N number of sampled result, in each sampled result, comprise the partial feature value of M eigenwert;
For each sampled result, k gaussian kernel is adopted to described sampled result, generates the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel;
After described N number of GMM submodel is sorted, obtain corresponding GMM model, and obtain described GMM model characteristic of correspondence vector, and utilize described proper vector to carry out feature detection.
Described feature extraction is carried out to video image, obtains the process of M eigenwert, specifically comprise:
Convergent-divergent process is carried out to described video image, obtains the video image of S different scale;
HOG Feature Descriptor is adopted to carry out feature extraction to each video image in the video image of a described S different scale respectively, to extract M1 eigenwert in the video image from each yardstick;
By M1 the eigenwert extracted respectively in the video image of a described S different scale, be combined into the characteristic set comprising M1*S eigenwert, to obtain M eigenwert, M=M1*S.
The process of described acquisition described GMM model characteristic of correspondence vector, specifically comprises:
The each gaussian kernel corresponding to described GMM model carries out parameter fitting, obtains center corresponding to each gaussian kernel and covariance;
The center corresponding to each gaussian kernel and covariance carry out differentiate process;
Described GMM model characteristic of correspondence vector is constructed according to the result of differentiate process.
Described method comprises further:
For each video image in a T to be detected video image, perform and feature extraction is carried out to video image, obtain M eigenwert, N sampling is carried out to M eigenwert, obtain N number of sampled result, the partial feature value of M eigenwert is comprised in each sampled result, for each sampled result, k gaussian kernel is adopted to described sampled result, generate the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel, after described N number of GMM submodel is sorted, obtain the process of corresponding GMM model, and obtain T GMM model;
After described T GMM model is sorted, obtain corresponding GMM reference model; For other video image to be detected outside a described T video image, feature extraction is carried out to other video image to be detected, utilize the eigenwert of current extraction and described GMM reference model to obtain the proper vector of other video image to be detected, and the proper vector of the video image utilizing other to be detected carry out feature detection.
Described method is applied in is undertaken in the system of feature detection by FisherVector mode;
The quantity k of described gaussian kernel is less than gaussian kernel quantity corresponding to a described M eigenwert.
The invention provides a kind of device of feature detection, for each video image to be detected, described device specifically comprises:
Extraction module, for carrying out feature extraction to video image, obtains M eigenwert;
Sampling module, for carrying out N sampling to a described M eigenwert, obtains N number of sampled result, comprises the partial feature value of a described M eigenwert in each sampled result;
Generation module, for for each sampled result, adopts k gaussian kernel to described sampled result, generates the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel;
Obtain module, after described N number of GMM submodel is sorted, obtain corresponding GMM model, and obtain described GMM model characteristic of correspondence vector;
Detection module, carries out feature detection for utilizing described proper vector.
Described extraction module, specifically for carrying out feature extraction to video image, obtaining, in the process of M eigenwert, carrying out convergent-divergent process to described video image, obtaining the video image of S different scale; HOG Feature Descriptor is adopted to carry out feature extraction to each video image in the video image of a described S different scale respectively, to extract M1 eigenwert in the video image from each yardstick; By M1 the eigenwert extracted respectively in the video image of a described S different scale, be combined into the characteristic set comprising M1*S eigenwert, to obtain M eigenwert, M=M1*S.
Described acquisition module, specifically in the process obtaining described GMM model characteristic of correspondence vector, each gaussian kernel corresponding to described GMM model carries out parameter fitting, obtains center corresponding to each gaussian kernel and covariance; The center corresponding to each gaussian kernel and covariance carry out differentiate process; Described GMM model characteristic of correspondence vector is constructed according to the result of differentiate process.
Described extraction module, also for for each video image in a T to be detected video image, carries out feature extraction to video image, obtains M eigenwert; Described sampling module, also for for each video image in a T to be detected video image, carries out N sampling to a corresponding M eigenwert, obtains N number of sampled result, comprise the partial feature value of a described M eigenwert in each sampled result; Described generation module, also for for each video image in a T to be detected video image, for each sampled result, adopts k gaussian kernel to described sampled result, generates the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel; Described acquisition module, also for for each video image in a T to be detected video image, after sorting, obtains corresponding GMM model, and obtains T GMM model described N number of GMM submodel;
Described acquisition module, also for after sorting to described T GMM model, obtains corresponding GMM reference model; Described extraction module, also for for other video image to be detected outside a described T video image, carries out feature extraction to other video image to be detected; Described acquisition module, the proper vector also for utilizing the eigenwert of current extraction and GMM reference model to obtain other video image to be detected; Described detection module, the proper vector also for the video image utilizing other to be detected carries out feature detection.
Described application of installation is being undertaken in the system of feature detection by FisherVector mode;
The quantity k of described gaussian kernel is less than gaussian kernel quantity corresponding to a described M eigenwert.
Based on technique scheme, in the embodiment of the present invention, for each video image to be detected, by carrying out feature extraction to video image, obtain M eigenwert, N sampling is carried out to M eigenwert, obtain N number of sampled result, the partial feature value of M eigenwert is comprised in each sampled result, for each sampled result, k gaussian kernel is adopted to sampled result, generate the GMM submodel that sampled result is corresponding, and obtain N number of GMM submodel, after N number of GMM submodel is sorted, obtain corresponding GMM model, and obtain GMM model characteristic of correspondence vector, and utilize proper vector to carry out feature detection.Above-mentioned feature detection mode is the mode of being carried out feature detection by FisherVector mode, not directly generate corresponding high dimensional feature vector based on original video image, its reliability is higher, and, above-mentioned feature detection mode is not directly adopt cascade system to generate high dimensional feature vector, can not produce feature inconsistent phenomenon when carrying out feature detection.Further, above-mentioned feature detection mode can use less k gaussian kernel to generate corresponding GMM submodel, the gaussian kernel quantity used in the generative process of GMM submodel is little, thus reduce the computation complexity of feature detection algorithm, improve calculated performance, improve Algorithm Convergence, and effectively can accelerate the generation of proper vector.
Accompanying drawing explanation
Fig. 1 is the process flow diagram of the method for feature detection in one embodiment of the present invention;
Fig. 2 is the hardware structure diagram of the image processing equipment in one embodiment of the present invention;
Fig. 3 is the structural drawing of the device of feature detection in one embodiment of the present invention.
Embodiment
For problems of the prior art, propose a kind of method of feature detection in the embodiment of the present invention, the method can be applied in is undertaken in the system of feature detection by FisherVector (Fei Sheer vector) mode.Wherein, FisherVector mode is a kind of feature extraction mode based on mixed Gauss model, go to simulate the distribution of local feature by adopting K gaussian kernel to video image, effectively can merge local feature, and to the change of video image, there is stronger reliability, thus FisherVector mode is a kind of effective feature coding mode.But, FisherVector mode is when generating feature vector, need to adopt more gaussian kernel to generate GMM (GaussianMixtureModel, gauss hybrid models) model, as usually needed to adopt K gaussian kernel, and the value of K is generally 256-512, because K value is comparatively large, the computation complexity therefore generating GMM model is higher, and computing velocity reduces, constringency performance reduces, and often needs the long period to produce the proper vector of video image.Based on above-mentioned discovery, as shown in Figure 1, for each video image to be detected, the method for the feature detection proposed in the embodiment of the present invention, specifically can comprise the following steps:
Step 101, carries out feature extraction to video image, obtains M eigenwert.
Wherein, by intensive sampling algorithm, video image can be divided into the identical grid of M size, adopt HOG (HistogramofOrientedGradient, histograms of oriented gradients) Feature Descriptor, respectively feature extraction is carried out to the video image in M grid, thus obtain M corresponding eigenwert.
In the embodiment of the present invention, feature extraction is carried out to video image, obtain the process of M eigenwert, specifically can include but not limited to as under type: convergent-divergent process is carried out to video image, obtain the video image of S different scale; HOG Feature Descriptor is adopted to carry out feature extraction to each video image in the video image of S different scale respectively, to extract M1 eigenwert in the video image from each yardstick; By M1 the eigenwert extracted respectively in the video image of S different scale, be combined into the characteristic set comprising M1*S eigenwert, to obtain M eigenwert, M=M1*S.
Such as, the video image 1 of original video image to be yardstick be 10*10, carries out convergent-divergent process to video image 1, obtains the video image 2 of 5*5, the video image 3 of 20*20, video image 1, video image 2 and video image 3 form the video image of 3 (S) different scales.Adopt HOG Feature Descriptor to carry out feature extraction to video image 1, from video image 1, extract M1 eigenwert; Adopt HOG Feature Descriptor to carry out feature extraction to video image 2, from video image 2, extract M1 eigenwert; Adopt HOG Feature Descriptor to carry out feature extraction to video image 3, from video image 1, extract M1 eigenwert.By M1 the eigenwert extracted respectively from 3 video images, be combined into the characteristic set comprising M1*S eigenwert, to obtain M (M1*S) eigenwert.
In the embodiment of the present invention, above-mentioned video image can be the video image comprising face.
Step 102, carries out N time (as 10 times) sampling to M eigenwert, obtains N number of sampled result.Wherein, the partial feature value in M eigenwert is comprised in each sampled result.
Such as, M eigenwert is eigenwert 1-eigenwert 100, carries out first time sampling, obtain sampled result 1, comprise eigenwert 1, eigenwert 2 in this sampled result 1 M eigenwert.Second time sampling is carried out to M eigenwert, obtains sampled result 2, in this sampled result 2, comprise eigenwert 3, eigenwert 5, eigenwert 6.Third time sampling is carried out to M eigenwert, obtains sampled result 3, in this sampled result 3, comprise eigenwert 2, eigenwert 8, eigenwert 10.4th sampling is carried out to M eigenwert, obtains sampled result 4, in this sampled result 4, comprise eigenwert 51, eigenwert 52, eigenwert 53.By that analogy, until carry out N sampling to M eigenwert, this sampling process repeats no more.
Wherein, when carrying out N sampling to M eigenwert, Bagging (bagging) algorithm can be adopted to carry out N stochastic sampling to M eigenwert, obtain N number of sampled result.
Wherein, in the process of M eigenwert being carried out to N sampling, all eigenwerts comprised in N number of sampled result can be ensured as far as possible, include each eigenwert in M eigenwert.
Wherein, randomly draw in the process of N number of sampled result at employing Bagging algorithm from M eigenwert, each eigenwert in M eigenwert is comprised in order to make all eigenwerts comprised in N number of sampled result, can adopt less sampling rate (when namely sampling at every turn, gather a small amount of eigenwert), and sampling repeatedly, the value as N can be more greatly.In this manner, the eigenwert causing comprising in N number of sampled result and M eigenwert there is certain deviation, namely all eigenwerts comprised in N number of sampled result, may can not comprise each eigenwert in M eigenwert, based on this, after obtaining M eigenwert, can according to 3 σ principles (i.e. 3 σ principles of the normal distribution) eigenwert that its large deviations of elimination is larger.
Step 103, for each sampled result, adopts k gaussian kernel to this sampled result, generates the GMM submodel that this sampled result is corresponding, and obtain N number of GMM submodel.
Wherein, the quantity k of gaussian kernel is k gaussian kernel corresponding to the eigenwert quantity that comprises in sampled result.
Wherein, the quantity k of gaussian kernel is less than gaussian kernel quantity K corresponding to M eigenwert.
Such as, comprise eigenwert 1, eigenwert 2 in sampled result 1, and two eigenwerts only may need configuration 1 gaussian kernel, therefore, 1 gaussian kernel can be utilized to generate the GMM submodel 1 of sampled result 1 correspondence.Comprise eigenwert 3, eigenwert 5, eigenwert 6 in sampled result 2, and three eigenwerts only may need configuration 2 gaussian kernel, therefore, 2 gaussian kernel can be utilized to generate the GMM submodel 2 of sampled result 2 correspondence.By that analogy, generate a corresponding GMM submodel for each sampled result, the quantity due to sampled result is N number of, therefore can obtain N number of GMM submodel.
Step 104, after sorting, obtains corresponding GMM model to N number of GMM submodel, and obtains this GMM model characteristic of correspondence vector, and utilizes this proper vector to carry out feature detection.
Wherein, after obtaining N number of GMM submodel, this N number of GMM submodel can also be combined into GMM model, this GMM model is the GMM model corresponding to M eigenwert.Wherein, can sort to N number of GMM submodel, be regarded as the approximate of unified higher-dimension GMM model.
Another kind of implementation is unlike this: after obtaining M eigenwert, utilizes K the gaussian kernel that eigenwert quantity M is corresponding, directly generate the GMM model corresponding to M eigenwert, and the quantity k of gaussian kernel is less than the quantity K of gaussian kernel.Such as, when eigenwert quantity M is 1000, the value of the quantity K of gaussian kernel is generally 256-512, and the value as K is 512.Because K value is comparatively large, the computation complexity therefore generating GMM model is higher, and computing velocity reduces, and constringency performance reduces, and often needs the long period to produce the proper vector of video image.And in the embodiment of the present invention, the value of the quantity k of gaussian kernel can be the numerical value such as 2,3, and k value is less, therefore the computation complexity generating GMM submodel is lower, computing velocity improves, and constringency performance improves, and does not need the long period to produce the proper vector of video image.In addition, by N number of GMM submodel is combined into GMM model, the GMM model that this is combined into is similar to the GMM model corresponding to M eigenwert directly generated.
Based on technique scheme, because face belongs to non-rigid targets, therefore when detecting face, needing to collect more information and being described, thus the discrimination of face can be improved.But the information of collection more more easily produces noise at most, therefore after collecting the Feature Descriptor of face, need to carry out dimensionality reduction operation to face characteristic, its effect is the discrimination increasing feature reducing characteristic dimension while.In order to find low dimension formulation corresponding to high dimensional feature, linear mapping can be utilized to be mapped in lower dimensional space by high dimensional feature, and any A dimensional vector can be mapped in B n-dimensional subspace n by FisherVector.Based on this, in the embodiment of the present invention, convergent-divergent process is carried out to video image, obtain the video image of S different scale, and each video image is divided into the identical grid of M1 size, adopt HOG Feature Descriptor, respectively feature extraction is carried out to the video image in M1 grid of each video image in the video image of S different scale, to extract M1 eigenwert in the video image from each yardstick; Afterwards, by M1 the eigenwert extracted respectively in the video image of S different scale, be combined into the characteristic set comprising M1*S eigenwert, to obtain M eigenwert.Afterwards, N sampling is carried out to M eigenwert, obtain N number of sampled result, for each sampled result, utilize k the gaussian kernel that the eigenwert quantity that comprises in sampled result is corresponding, generate the GMM submodel that this sampled result is corresponding, obtain N number of GMM submodel, this N number of GMM submodel is combined into GMM model.Wherein, the gaussian kernel quantity that aforesaid way uses is the gaussian kernel quantity sum that N number of sampled result uses.
In the embodiment of the present invention, obtain the process of this GMM model characteristic of correspondence vector, specifically can include but not limited to as under type: each gaussian kernel corresponding to this GMM model carries out parameter fitting, obtains center corresponding to each gaussian kernel and covariance; The center corresponding to each gaussian kernel and covariance carry out differentiate process; This GMM model characteristic of correspondence vector is constructed according to the result of differentiate process.
Wherein, each gaussian kernel corresponding to this GMM model respectively carries out parameter fitting, obtain the process of center corresponding to each gaussian kernel and covariance, specifically can include but not limited to as under type: adopt EM (ExpectationMaximization, greatest hope) each gaussian kernel that algorithm is corresponding to GMM model carries out parameter fitting (namely carrying out parameter fitting to the GMM model of each gaussian kernel), obtain center corresponding to each gaussian kernel and covariance, this process repeats no longer in detail.
Wherein, the center corresponding to each gaussian kernel and covariance carry out the process of differentiate process, specifically can include but not limited to as under type: utilize following differentiate formula respectively, and the center corresponding to each gaussian kernel and covariance carry out differentiate process.
Φ k ( 1 ) = 1 N w k Σ p = 1 N α ( k ) ( x p - μ k σ k ) ;
Φ k ( 2 ) = 1 N - W k Σ p = 1 N α p ( k ) ( ( x p - μ k ) 2 σ k 2 - 1 ) .
Wherein, be respectively the weight of GMM model, center and covariance matrix.
Wherein, the process of this GMM model characteristic of correspondence vector is constructed according to the result of differentiate process, specifically can include but not limited to as under type: when differentiate process is carried out at center corresponding to each gaussian kernel corresponding to GMM model and covariance, can obtain the proper vector of a 2k*d dimension, the proper vector that this 2k*d ties up is this GMM model characteristic of correspondence vector.Wherein, k represents the gaussian kernel number of GMM model, and d represents the dimension of HOG feature.Wherein, this GMM model characteristic of correspondence vector is higher-dimension FisherVector proper vector, using the proper vector of this proper vector as video image (comprising the video image of face), and this proper vector can be utilized to carry out feature detection, as subsequent operations such as recognitions of face.
Based on technique scheme, in the embodiment of the present invention, for each video image to be detected, by carrying out feature extraction to video image, obtain M eigenwert, N sampling is carried out to M eigenwert, obtain N number of sampled result, the partial feature value of M eigenwert is comprised in each sampled result, for each sampled result, k gaussian kernel is adopted to sampled result, generate the GMM submodel that sampled result is corresponding, and obtain N number of GMM submodel, after N number of GMM submodel is sorted, obtain corresponding GMM model, and obtain GMM model characteristic of correspondence vector, and utilize proper vector to carry out feature detection.
Above-mentioned feature detection mode is the mode of being carried out feature detection by FisherVector mode, not directly generate corresponding high dimensional feature vector based on original video image, its reliability is higher, and, above-mentioned feature detection mode is not directly adopt cascade system to generate high dimensional feature vector, can not produce feature inconsistent phenomenon when carrying out feature detection.Further, above-mentioned feature detection mode can also use less k gaussian kernel to generate corresponding GMM submodel, the gaussian kernel quantity used in the generative process of GMM submodel is little, thus reduce the computation complexity of feature detection algorithm, and can calculated performance be improved, improve Algorithm Convergence, and effectively can accelerate the generation of proper vector.
Compared with using the FisherVector method of K gaussian kernel, use the EM convergence algorithm of the FisherVector method of K gaussian kernel to affect by original state comparatively large, speed of convergence is slower.In the embodiment of the present invention, by sampling to M eigenwert, and use the less GMM submodel of Gauss's check figure to estimate to the eigenwert after sampling, by carrying out multiple repairing weld to M eigenwert, thus abundant parameter can be estimated, and due to each estimated parameter less, therefore, EM convergence algorithm can be made to be restrained under less iterations, thus there is estimating speed faster.
In order to improve the speed of feature extraction further, in the embodiment of the present invention, for each video image in a T to be detected video image, perform and feature extraction is carried out to video image, obtain M eigenwert, N sampling is carried out to M eigenwert, obtain N number of sampled result, the partial feature value of M eigenwert is comprised in each sampled result, for each sampled result, k gaussian kernel is adopted to this sampled result, generate the GMM submodel that this sampled result is corresponding, and obtain N number of GMM submodel, after N number of GMM submodel is sorted, obtain the process (namely step 101-step 103 being performed to each video image in a T to be detected video image) of corresponding GMM model, and obtain T GMM model.Afterwards, after T GMM model is sorted, obtain corresponding GMM reference model.Based on this, for other video image to be detected outside T video image, directly feature extraction is carried out to other video image to be detected, and utilize eigenwert and the GMM reference model of current extraction, obtain the proper vector of other video image to be detected, and the proper vector of the video image utilizing other to be detected carries out feature detection.
Wherein, after obtaining GMM model corresponding to each video image in T video image, can using T GMM model obtaining as T GMM submodel, and after T GMM submodel is sorted, obtain the GMM model that this T GMM submodel is corresponding, and using this GMM model as GMM reference model.
Wherein, for other video image (facial image), by HOG feature (eigenwert) and the GMM reference model of other video image, FisherVector proper vector corresponding to other video image can be generated to obtain, and utilize the proper vector of other video image to carry out feature detection.HOG feature corresponding to different facial image is different, and therefore, different facial image can produce different proper vectors.
Based on above-mentioned implementation, do not need to perform for each video image the process generating GMM model, significantly reduce computation complexity, there is lower computation complexity, and improve recognition rate.
Based on the inventive concept same with said method, a kind of device of feature detection is additionally provided in the embodiment of the present invention, the device of this feature detection can be applied in image processing equipment, the device of this feature detection can pass through software simulating, also can be realized by the mode of hardware or software and hardware combining.For software simulating, as the device on a logical meaning, be the processor of the image processing equipment by its place, computer program instructions corresponding in nonvolatile memory read in internal memory to run and formed.Say from hardware view, as shown in Figure 2, for a kind of hardware structure diagram of the image processing equipment at the device place of the feature detection of the present invention's proposition, except the processor shown in Fig. 2, network interface, internal memory and nonvolatile memory, image processing equipment can also comprise other hardware, as the forwarding chip etc. of responsible process message; From hardware configuration, this image processing equipment may be also distributed apparatus, may comprise multiple interface card, to carry out the expansion of Message processing at hardware view.
As shown in Figure 3, be the structural drawing of the device of the feature detection of the present invention's proposition, the device of this feature detection can be applied in image processing equipment, and for each video image to be detected, the device of described feature detection specifically comprises:
Extraction module 11, for carrying out feature extraction to video image, obtains M eigenwert;
Sampling module 12, for carrying out N sampling to a described M eigenwert, obtains N number of sampled result, comprises the partial feature value of a described M eigenwert in each sampled result;
Generation module 13, for for each sampled result, adopts k gaussian kernel to described sampled result, generates the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel;
Obtain module 14, after described N number of GMM submodel is sorted, obtain corresponding GMM model, and obtain described GMM model characteristic of correspondence vector;
Detection module 15, carries out feature detection for utilizing described proper vector.
Described extraction module 11, specifically for carrying out feature extraction to video image, obtaining, in the process of M eigenwert, carrying out convergent-divergent process to described video image, obtaining the video image of S different scale; HOG Feature Descriptor is adopted to carry out feature extraction to each video image in the video image of a described S different scale respectively, to extract M1 eigenwert in the video image from each yardstick; By M1 the eigenwert extracted respectively in the video image of a described S different scale, be combined into the characteristic set comprising M1*S eigenwert, to obtain M eigenwert, M=M1*S.
In the embodiment of the present invention, described acquisition module 14, specifically in the process obtaining described GMM model characteristic of correspondence vector, each gaussian kernel corresponding to described GMM model carries out parameter fitting, obtains center corresponding to each gaussian kernel and covariance; The center corresponding to each gaussian kernel and covariance carry out differentiate process; Described GMM model characteristic of correspondence vector is constructed according to the result of differentiate process.
Described extraction module 11, also for for each video image in a T to be detected video image, carries out feature extraction to video image, obtains M eigenwert; Described sampling module 12, also for for each video image in a T to be detected video image, carries out N sampling to a corresponding M eigenwert, obtains N number of sampled result, comprise the partial feature value of a described M eigenwert in each sampled result; Described generation module 13, also for for each video image in a T to be detected video image, for each sampled result, adopts k gaussian kernel to described sampled result, generates the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel; Described acquisition module 14, also for for each video image in a T to be detected video image, after sorting, obtains corresponding GMM model, and obtains T GMM model described N number of GMM submodel;
Described acquisition module 14, also for after sorting to described T GMM model, obtains corresponding GMM reference model; Described extraction module 11, also for for other video image to be detected outside a described T video image, carries out feature extraction to other video image to be detected; Described acquisition module 14, the proper vector also for utilizing the eigenwert of current extraction and GMM reference model to obtain other video image to be detected; Described detection module 15, the proper vector also for the video image utilizing other to be detected carries out feature detection.
Described application of installation is being undertaken in the system of feature detection by FisherVector mode;
The quantity k of described gaussian kernel is less than gaussian kernel quantity corresponding to a described M eigenwert.
Wherein, the modules of apparatus of the present invention can be integrated in one, and also can be separated deployment.Above-mentioned module can merge into a module, also can split into multiple submodule further.
Through the above description of the embodiments, those skilled in the art can be well understood to the mode that the present invention can add required general hardware platform by software and realize, and can certainly pass through hardware, but in a lot of situation, the former is better embodiment.Based on such understanding, technical scheme of the present invention can embody with the form of software product the part that prior art contributes in essence in other words, this computer software product is stored in a storage medium, comprising some instructions in order to make a computer equipment (can be personal computer, server, or the network equipment etc.) perform method described in each embodiment of the present invention.It will be appreciated by those skilled in the art that accompanying drawing is the schematic diagram of a preferred embodiment, the module in accompanying drawing or flow process might not be that enforcement the present invention is necessary.
It will be appreciated by those skilled in the art that the module in the device in embodiment can carry out being distributed in the device of embodiment according to embodiment description, also can carry out respective change and be arranged in the one or more devices being different from the present embodiment.The module of above-described embodiment can merge into a module, also can split into multiple submodule further.The invention described above embodiment sequence number, just to describing, does not represent the quality of embodiment.
Be only several specific embodiment of the present invention above, but the present invention is not limited thereto, the changes that any person skilled in the art can think of all should fall into protection scope of the present invention.

Claims (10)

1. a method for feature detection, is characterized in that, for each video image to be detected, said method comprising the steps of:
Feature extraction is carried out to video image, obtains M eigenwert; N sampling is carried out to M eigenwert, obtains N number of sampled result, in each sampled result, comprise the partial feature value of M eigenwert;
For each sampled result, k gaussian kernel is adopted to described sampled result, generates the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel;
After described N number of GMM submodel is sorted, obtain corresponding GMM model, and obtain described GMM model characteristic of correspondence vector, and utilize described proper vector to carry out feature detection.
2. method according to claim 1, is characterized in that, describedly carries out feature extraction to video image, obtains the process of M eigenwert, specifically comprises:
Convergent-divergent process is carried out to described video image, obtains the video image of S different scale;
HOG Feature Descriptor is adopted to carry out feature extraction to each video image in the video image of a described S different scale respectively, to extract M1 eigenwert in the video image from each yardstick;
By M1 the eigenwert extracted respectively in the video image of a described S different scale, be combined into the characteristic set comprising M1*S eigenwert, to obtain M eigenwert, M=M1*S.
3. method according to claim 1, is characterized in that, the process of described acquisition described GMM model characteristic of correspondence vector, specifically comprises:
The each gaussian kernel corresponding to described GMM model carries out parameter fitting, obtains center corresponding to each gaussian kernel and covariance;
The center corresponding to each gaussian kernel and covariance carry out differentiate process;
Described GMM model characteristic of correspondence vector is constructed according to the result of differentiate process.
4. method according to claim 1, is characterized in that, described method comprises further:
For each video image in a T to be detected video image, perform and feature extraction is carried out to video image, obtain M eigenwert, N sampling is carried out to M eigenwert, obtain N number of sampled result, the partial feature value of M eigenwert is comprised in each sampled result, for each sampled result, k gaussian kernel is adopted to described sampled result, generate the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel, after described N number of GMM submodel is sorted, obtain the process of corresponding GMM model, and obtain T GMM model;
After described T GMM model is sorted, obtain corresponding GMM reference model;
For other video image to be detected outside a described T video image, feature extraction is carried out to other video image to be detected, utilize the eigenwert of current extraction and described GMM reference model to obtain the proper vector of other video image to be detected, and the proper vector of the video image utilizing other to be detected carry out feature detection.
5. the method according to any one of claim 1-4, is characterized in that, described method is applied in is undertaken in the system of feature detection by FisherVector mode;
The quantity k of described gaussian kernel is less than gaussian kernel quantity corresponding to a described M eigenwert.
6. a device for feature detection, is characterized in that, for each video image to be detected, described device specifically comprises:
Extraction module, for carrying out feature extraction to video image, obtains M eigenwert;
Sampling module, for carrying out N sampling to a described M eigenwert, obtains N number of sampled result, comprises the partial feature value of a described M eigenwert in each sampled result;
Generation module, for for each sampled result, adopts k gaussian kernel to described sampled result, generates the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel;
Obtain module, after described N number of GMM submodel is sorted, obtain corresponding GMM model, and obtain described GMM model characteristic of correspondence vector;
Detection module, carries out feature detection for utilizing described proper vector.
7. device according to claim 6, is characterized in that,
Described extraction module, specifically for carrying out feature extraction to video image, obtaining, in the process of M eigenwert, carrying out convergent-divergent process to described video image, obtaining the video image of S different scale; HOG Feature Descriptor is adopted to carry out feature extraction to each video image in the video image of a described S different scale respectively, to extract M1 eigenwert in the video image from each yardstick; By M1 the eigenwert extracted respectively in the video image of a described S different scale, be combined into the characteristic set comprising M1*S eigenwert, to obtain M eigenwert, M=M1*S.
8. device according to claim 6, is characterized in that,
Described acquisition module, specifically in the process obtaining described GMM model characteristic of correspondence vector, each gaussian kernel corresponding to described GMM model carries out parameter fitting, obtains center corresponding to each gaussian kernel and covariance; The center corresponding to each gaussian kernel and covariance carry out differentiate process; Described GMM model characteristic of correspondence vector is constructed according to the result of differentiate process.
9. device according to claim 6, is characterized in that,
Described extraction module, also for for each video image in a T to be detected video image, carries out feature extraction to video image, obtains M eigenwert; Described sampling module, also for for each video image in a T to be detected video image, carries out N sampling to a corresponding M eigenwert, obtains N number of sampled result, comprise the partial feature value of a described M eigenwert in each sampled result; Described generation module, also for for each video image in a T to be detected video image, for each sampled result, adopts k gaussian kernel to described sampled result, generates the GMM submodel that described sampled result is corresponding, and obtain N number of GMM submodel; Described acquisition module, also for for each video image in a T to be detected video image, after sorting, obtains corresponding GMM model, and obtains T GMM model described N number of GMM submodel;
Described acquisition module, also for after sorting to described T GMM model, obtains corresponding GMM reference model; Described extraction module, also for for other video image to be detected outside a described T video image, carries out feature extraction to other video image to be detected; Described acquisition module, the proper vector also for utilizing the eigenwert of current extraction and GMM reference model to obtain other video image to be detected; Described detection module, the proper vector also for the video image utilizing other to be detected carries out feature detection.
10. the device according to any one of claim 6-9, is characterized in that, described application of installation is being undertaken in the system of feature detection by FisherVector mode;
The quantity k of described gaussian kernel is less than gaussian kernel quantity corresponding to a described M eigenwert.
CN201510727485.9A 2015-10-30 2015-10-30 Feature detection method and device Active CN105354579B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201510727485.9A CN105354579B (en) 2015-10-30 2015-10-30 Feature detection method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201510727485.9A CN105354579B (en) 2015-10-30 2015-10-30 Feature detection method and device

Publications (2)

Publication Number Publication Date
CN105354579A true CN105354579A (en) 2016-02-24
CN105354579B CN105354579B (en) 2020-07-28

Family

ID=55330546

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201510727485.9A Active CN105354579B (en) 2015-10-30 2015-10-30 Feature detection method and device

Country Status (1)

Country Link
CN (1) CN105354579B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019201042A1 (en) * 2018-04-16 2019-10-24 腾讯科技（深圳）有限公司 Image object recognition method and device, storage medium, and electronic device

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070140550A1 (en) * 2005-12-20 2007-06-21 General Instrument Corporation Method and apparatus for performing object detection
CN101236599A (en) * 2007-12-29 2008-08-06 浙江工业大学 Human face recognition detection device based on multi- video camera information integration
CN102087707A (en) * 2009-12-03 2011-06-08 索尼株式会社 Image processing equipment and image processing method
US20120184840A1 (en) * 2009-04-07 2012-07-19 Kayvan Najarian Automated Measurement of Brain Injury Indices Using Brain CT Images, Injury Data, and Machine Learning
CN103927561A (en) * 2014-04-29 2014-07-16 东南大学 Behavior recognition method based on probability fusion and dimensionality reduction technology
CN104134059A (en) * 2014-07-25 2014-11-05 西安电子科技大学 Color-information-maintaining objectionable image detection method under deformation sensitive organ models
CN104820843A (en) * 2015-05-29 2015-08-05 常熟苏大低碳应用技术研究院有限公司 Method for marking picture semantics based on Gauss mixture model

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070140550A1 (en) * 2005-12-20 2007-06-21 General Instrument Corporation Method and apparatus for performing object detection
CN101236599A (en) * 2007-12-29 2008-08-06 浙江工业大学 Human face recognition detection device based on multi- video camera information integration
US20120184840A1 (en) * 2009-04-07 2012-07-19 Kayvan Najarian Automated Measurement of Brain Injury Indices Using Brain CT Images, Injury Data, and Machine Learning
CN102087707A (en) * 2009-12-03 2011-06-08 索尼株式会社 Image processing equipment and image processing method
CN103927561A (en) * 2014-04-29 2014-07-16 东南大学 Behavior recognition method based on probability fusion and dimensionality reduction technology
CN104134059A (en) * 2014-07-25 2014-11-05 西安电子科技大学 Color-information-maintaining objectionable image detection method under deformation sensitive organ models
CN104820843A (en) * 2015-05-29 2015-08-05 常熟苏大低碳应用技术研究院有限公司 Method for marking picture semantics based on Gauss mixture model

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
刘轩等: "基于数字图像处理的人员运动参数提取方法研究", 《火灾科学》 *
方晶: "语音身份和内容同时识别技术及其应用研究", 《中国优秀硕士学位论文全文数据库》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019201042A1 (en) * 2018-04-16 2019-10-24 腾讯科技（深圳）有限公司 Image object recognition method and device, storage medium, and electronic device
US11227182B2 (en) * 2018-04-16 2022-01-18 Tencent Technology (Shenzhen) Company Limited Method, apparatus, and storage medium for recognizing image object

Also Published As

Publication number Publication date
CN105354579B (en) 2020-07-28

Similar Documents

Publication Publication Date Title
Ohn-Bar et al. Joint angles similarities and HOG2 for action recognition
Zhi et al. LightNet: A Lightweight 3D Convolutional Neural Network for Real-Time 3D Object Recognition.
Rungsuptaweekoon et al. Evaluating the power efficiency of deep learning inference on embedded GPU systems
Gong et al. Object detection based on improved YOLOv3-tiny
CN105320764A (en) 3D model retrieval method and 3D model retrieval apparatus based on slow increment features
CN104809478B (en) A kind of image block method and device towards extensive three-dimensional reconstruction
Suzuki et al. Multiple-clue face detection algorithm using edge-based feature vectors
Yang et al. Towards improving classification power for one-shot object detection
CN106203508A (en) A kind of image classification method based on Hadoop platform
Iwashita et al. Gait recognition robust to speed transition using mutual subspace method
Zhou et al. A pipeline architecture for traffic sign classification on an FPGA
Chen et al. Action recognition using lie algebrized gaussians over dense local spatio-temporal features
CN117375896A (en) Intrusion detection method and system based on multi-scale space-time feature residual fusion
Kim et al. Image recognition accelerator design using in-memory processing
CN105354579A (en) Feature detection method and apparatus
Adiono et al. HOG-AdaBoost implementation for human detection employing FPGA ALTERA DE2-115
CN113378722B (en) Behavior identification method and system based on 3D convolution and multilevel semantic information fusion
Ji et al. Deep age estimation model stabilization from images to videos
Tan et al. Divide to attend: A multiple receptive field attention module for object detection in remote sensing images
WANG et al. GAN-SR Anomaly Detection Model Based on Imbalanced Data
Shekade et al. Vehicle Classification in Traffic Surveillance System using YOLOv3 Model
CN112348060A (en) Classification vector generation method and device, computer equipment and storage medium
Xu et al. Feature-enhanced occlusion perception object detection for smart cities
Lai et al. Communication-aware face detection using noc architecture
Jabshetti et al. Object detection using Regionlet transform

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant