CN109241981A - A kind of characteristic detection method based on sparse coding - Google Patents
A kind of characteristic detection method based on sparse coding Download PDFInfo
- Publication number
- CN109241981A CN109241981A CN201811021166.6A CN201811021166A CN109241981A CN 109241981 A CN109241981 A CN 109241981A CN 201811021166 A CN201811021166 A CN 201811021166A CN 109241981 A CN109241981 A CN 109241981A
- Authority
- CN
- China
- Prior art keywords
- sparse
- dictionary
- image block
- point
- local
- 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
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
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2136—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on sparsity criteria, e.g. with an overcomplete basis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
In order to improve the accuracy of target following, the present invention provides a kind of characteristic detection method based on sparse coding, belongs to the target following technical field of computer vision field.The present invention includes: S1, using extracting local feature region based on the FAST Corner Detection Algorithm of local threshold, and calculates the Local gradient direction of characteristic point, using characteristic point as center sampled images block, as training sample；S2, it training is grouped to dictionary element according to the Local gradient direction of characteristic point obtained complete dictionary；S3, rarefaction representation is carried out to test sample using obtained dictionary, then to image block piecemeal, constructs sparse features, target following is realized according to the detector of the sparse features.The present invention utilizes the sparse features of sparse coding learning objective, improves the accuracy and robustness of target following.The dictionary element of different grouping is respectively trained according to the Local gradient direction of image block, to reflect the local direction information of image block, improves the accuracy of target following.
Description
Technical field
The present invention relates to a kind of target signature detection methods, in particular to a kind of sparse come learning objective based on sparse coding
The characteristic detection method of feature belongs to the target following technical field of computer vision field.
Background technique
In recent years, having been achieved for biggish progress for the algorithm of general target tracking.Tracking box based on detection
Frame (tracking-by-detection) is led by combining image detection and existing tracking technique in target following
Domain achieves a series of successes.The high efficiency and high accuracy for having benefited from detection algorithm, the tracking frame based on detection often compare
The existing track algorithm based on estimation has better accuracy and efficiency.Existing feature detection algorithm contains abundant
Local feature information, sufficiently target can be indicated and have lesser calculation amount, facilitate improve detection algorithm
Performance.Avidan of Mitsubishi Electric research laboratory et al. constructs spy using local direction histogram and RGB color
Vector is levied, proposes a kind of track algorithm based on detection with low computation complexity at first.Kalal of Surrey Roehampton et al.
Propose TLD (Tracking-Learning-Detection, tracking-study-detection) track algorithm.They are dedicated to LBP
(Local Binary Pattern, local binary patterns) feature is applied in tracking, and online updating during tracking
Classifier is to adapt to the variation of target appearance.The a collection of track algorithm based on correlation filtering, feature inspection are emerged recent years
Surveying device equally has important influence to the accuracy of entire track algorithm.Oxonian Henriques et al. proposes one
CSK (Exploiting the Circulant Structure of Tracking-by- of the kind based on original gradation feature
Detection with Kernels, utilizes the loop structure of kernel function detecting and tracking) algorithm, algorithm was expanded to again later more
In gradient orientation histogram (Histogram of Oriented Gradients, HOG) feature space of dimension, multi-pass is constructed
Road correlation filter proposes famous KCF in turn, and (correlation filtering of Kernel Correlation Filter, coring are calculated
Method).KCF classifies to target and background by solving ridge regression problem, and the weight matrix acquired is correlation filter.Core
Thought is that specimen sample is carried out by the way of circulating sampling, obtains circulation sample matrix, constructs training sample, and then pass through Fu
In leaf transformation in frequency domain rapid solving correlation filter.When being predicted, circulating sampling is first carried out, utilizes the phase being previously obtained
The similarity degree that filter calculates new test sample and target is closed, optimal test sample is finally obtained.KCF in robustness and
All there is preferable performance in efficiency, become the work of the foundation of later correlation filtering and depth track algorithm.It is above-mentioned this
The common ground of a little algorithms is will to track process as two classification problems, target appearance is indicated by feature detection, then
Foreground target is detected from background by one two classifier of training.
For the track algorithm based on detection, clarification of objective detection is most basic treatment process, is directly affected entire
The treatment processes such as the subsequent classification of tracking system and tracking.Therefore, how accurately and efficiently target to be indicated always
It is the emphasis of research.However, existing feature detection algorithm is often more effective only for the target of particular category, or limitation
In color, shape, edge, texture, the characteristic information in a certain respect such as local gray level.For example, Haar is field of face identification ratio
More effective detection algorithm can not effectively describe the color of target, the information such as texture.In addition, these features require craft
The design for carrying out a series of complex, is not easy to expand in other application scenarios.For this purpose, researcher, which begins one's study, utilizes nerve net
The representation of network learning characteristic from training sample, and huge success is achieved in image detection and classification field.By
The inspiration of image detection, the tracking of the track algorithm based on deep learning, i.e. depth become target tracking domain in recent years new
Hot spot.Held of Stamford et al. proposes a kind of high frame per second target tracking algorism GOTURN neural network based at first
(Generic Object Tracking Using Regression Networks).GOTURN is estimated using Recurrent networks
The motion conditions of target, then in forecast image target position, rather than clarification of objective table is extracted using neural network
Show and then classifies.In order to carry out feature detection, Oxford University using the powerful information extraction ability of Classification Neural
Bertinetto et al. a convolutional network is obtained as feature extractor by off-line training, then carry out feature detection.
On the basis of Bertinetto et al. work, correlation filter is introduced into neural network by Oxonian Valmadre et al.
In, CFNet (End-to-ena is proposed as one layer of convolutional layer therein to calculate the correlation of sample and training dataset
representation learning for Correlation Filter based tracking).By correlation filtering
Device, CFNet are significantly improved in accuracy.In addition to depth track algorithm end to end, the Danelljan of University of Linkoping
Et al. using neural network as an Analysis On Multi-scale Features detector, propose ECO (Efficient Convolution
Operators, effective convolution operator) algorithm.ECO utilizes VGG (VisualGeometryGroup, visual geometric group) classification net
The first layer and layer 5 of network come extract respectively target part and macroscopic view feature, then with these difference abstraction hierarchies feature
The convolution filter of training frequency domain, the correlation of test sample and training set is calculated by convolution algorithm, to be inferred to most
Good target position.Different from depth track algorithm end to end, ECO is utilized what deep learning learnt in classification task
Powerful character representation ability, then updates the parameter of convolution filter during on-line training according to depth characteristic, fits
Answer the cosmetic variation of target.Therefore, ECO is a kind of track algorithm similar to correlation filtering, and convolution filter therein is similar
In the correlation filter of KCF.This correlation filtering has good compatibility for different feature extraction algorithms,
It also can be very good tracking effect in HOG feature.But neural network model generally requires time-consuming off-line training process.
Summary of the invention
In order to improve the accuracy of target following, the present invention provides a kind of characteristic detection method based on sparse coding.
Characteristic detection method based on sparse coding of the invention, which comprises
S1, using extracting local feature region based on the FAST Corner Detection Algorithm of local threshold, and calculate characteristic point
Local gradient direction, using characteristic point as center sampled images block, as training sample；
S2, it training is grouped to dictionary element according to the Local gradient direction of characteristic point obtained complete dictionary；
S3, sparse spy is constructed then to image block piecemeal to test sample progress rarefaction representation using obtained dictionary
Sign realizes target following according to the detector of the sparse features.
Preferably, in S1, the local feature region of former frame samplings is extracted, as training sample, is adopted in S3 with present frame
Sample is as test sample.
Preferably, the S1 includes:
S11, image to be detected is switched to grayscale image, measuring point c to be checked is located at circle center, and surrounding point is located at circle void
On line, the gray value of measuring point to be checked is Ic, the gray value of surrounding point is Ic→p, the threshold value compared for gray value is T, such as
It is more than threshold value T that fruit, which has the difference of N number of gray value continuously put and measuring point to be checked, then measuring point to be checked is a characteristic point；Its
Middle threshold value T:
Wherein, C indicates the number of the point in 3 × 3 neighborhood of measuring point to be checked, IiIt is the gray scale put in 3 × 3 neighborhood of measuring point to be checked
Value, a indicate threshold factor, σ3×3Indicate the variance of gray value in neighborhood；Reaction
Be confidence level to threshold value, value range is [0,1]；
S12, the gradient direction for calculating characteristic point:
△Ix=Ix+1-Ix-1
△Iy=Iy+1-Iy-1
Wherein, △ IxIndicate lateral gray value difference, △ IyIndicate longitudinal difference, ori indicates gradient direction；
Gradient orientation histogram at S13, statistical nature point, by gradient direction discretization to 0 °, 45 °, 90 °, 135 ° 4
On section, the image block of 4 Local gradient directions of characteristic point is obtained.
Preferably, the S2 includes:
4 sub- dictionaries, trained 4 sub- dictionary groups is respectively trained using the image block of 4 different Local gradient directions
At excessively complete dictionary D；
Dictionary learning problem representation are as follows:
Wherein, Y=[y1,…,yk] it is the matrix that k training sample forms, each column vector ykIt is arranged by image block according to column
The mode conversion dimension of sequence obtains, X=[x1,…,xk] it is sparse coefficient matrix,Be to be solved it is excessively complete
Standby dictionary；K refers to degree of rarefication, i.e., each column vector x in sparse coefficient matrixiThe number of middle nonzero element, | | xi||0Indicate 0 model
Number；
The element in dictionary D is updated using based on the decline of the gradient of momentum:
Wherein, diIndicate the i-th column vector of dictionary D,Indicate sparse coefficient matrix the i-th row nonzero element composition row to
Amount,It is that dictionary element d can be used in matrix YiThe matrix for the training sample composition being indicated, it is non-with the i-th row of sparse matrix
Neutral element is corresponding,Indicate F norm；
Increment of sample vtUpdate method iteratively solves optimal dictionary element d, and then obtains complete dictionary D:
Wherein, v0It is the initial value of incremental update, vtIt is that intermediate variable is used for diIt is updated.μ is super on [0,1]
Parameter,
Preferably, the S3 includes:
S31, image block is acquired at each pixel of test sample image；
S32, rarefaction representation is carried out to image block by orthogonal matching pursuit OMP algorithm using obtained complete dictionary, obtained
Obtain the sparse coefficient of pixel in image block；
S33, to image block piecemeal:
Image block is divided into 4 × 4 small cube, 8 × 8 image block each in this way is made of 44 × 4 small cubes, and 4
Small cube regards the small cube of 8 × 8 image block upper lefts, lower-left, upper right, bottom right as respectively；
S34, according to the position of each pixel, the sparse coefficient of the pixel projected to using bilinear interpolation around
4 small cubes on, the corresponding sparse coefficient of 4 small cubes is obtained according to the weight that projects on 4 small cubes；
4 small cubes around described include that small cube where the pixel is outer and adjacent with the small cube three images
The nearest small cube of distance in block, 8 × 8 image blocks of 4 small cubes composition；
Project to the weight on 4 small cubes:
wcell1=dx2 × dy2
wcell2=dx1 × dy2
wcell3=dx2 × dy1
wcell4=dx1 × dy1
Wherein, dx1 and dx2 is left and right lateral distance of the pixel from four nearest small cube centers, and dy1 and dy2 are
Up and down fore-and-aft distance of the pixel from four nearest small cube centers；
S35, the feature vector that each small cube is constructed according to the corresponding sparse coefficient of 4 small cubes that S34 is obtained, the spy
Levying vector is three-dimensional eigenmatrix, and longitudinal dimension of the matrix normalizes to obtain final sparse features；
S36, target following is realized according to the detector of the sparse features.
Above-mentioned technical characteristic may be combined in various suitable ways or be substituted by equivalent technical characteristic, as long as can reach
To the purpose of the present invention.
The beneficial effects of the present invention are the present invention obtains dictionary by trained from sample data, utilizes sparse coding
The sparse features of learning objective solve existing feature detection in target following and need to carry out many and diverse hand-designed for target
The problem of, improve the accuracy and robustness of target following.When updating dictionary element, according to the Local gradient direction of image block
The dictionary element of different grouping is respectively trained, to reflect the local direction information of image block, improves the expression ability of feature detection,
And then improve the accuracy of target following.
Detailed description of the invention
Fig. 1 is the principle of the present invention schematic diagram；
Fig. 2 is sparse features projection process；
Fig. 3 is precision curve of the present embodiment on OTB100 test set, and ordinate Precision indicates precision, horizontal
Coordinate Threshold indicates threshold value；
Fig. 4 is success rate curve of the present embodiment on OTB100 test set, and ordinate Successrate is indicated successfully
Rate, abscissa Overlap threshold indicate anti-eclipse threshold；
Fig. 5 is that present embodiment background on OTB100 test set obscures the precision curve under scene, ordinate
Precision indicates precision, and abscissa Threshold indicates threshold value；
Fig. 6 is that present embodiment background on OTB100 test set obscures the success rate curve under scene, ordinate
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Fig. 7 is precision curve of the present embodiment on OTB100 test set under deformation scene, ordinate Precision table
Show precision, abscissa Threshold indicates threshold value；
Fig. 8 is success rate curve of the present embodiment on OTB100 test set under deformation scene, ordinate
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Fig. 9 is precision curve of the present embodiment on OTB100 test set under fast motion scenes, ordinate
Precision indicates precision, and abscissa Threshold indicates threshold value；
Figure 10 is success rate curve of the present embodiment on OTB100 test set under fast motion scenes, ordinate
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Figure 11 is precision curve of the present embodiment on OTB100 test set under illumination scene change, ordinate
Precision indicates precision, and abscissa Threshold indicates threshold value；
Figure 12 is success rate curve of the present embodiment on OTB100 test set under illumination scene change, ordinate
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Figure 13 is precision curve of the present embodiment on OTB100 test set under plane internal rotation scene, ordinate
Precision indicates precision, and abscissa Threshold indicates threshold value；
Figure 14 is success rate curve of the present embodiment on OTB100 test set under plane internal rotation scene, ordinate
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Figure 15 is precision curve of the present embodiment on OTB100 test set under low resolution scene, ordinate
Precision indicates precision, and abscissa Threshold indicates threshold value；
Figure 16 is success rate curve of the present embodiment on OTB100 test set under low resolution scene, ordinate
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Figure 17 is precision curve of the present embodiment on OTB100 test set under motion blur scene, ordinate
Precision indicates precision, and abscissa Threshold indicates threshold value；
Figure 18 is success rate curve of the present embodiment on OTB100 test set under motion blur scene, ordinate
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Figure 19 is that present embodiment blocks the precision curve under scene, ordinate Precision on OTB100 test set
Indicate precision, abscissa Threshold indicates threshold value；
Figure 20 is that present embodiment blocks the success rate curve under scene, ordinate on OTB100 test set
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Figure 21 is precision curve of the present embodiment on OTB100 test set under three-dimensional rotation scene, ordinate
Precision indicates precision, and abscissa Threshold indicates threshold value；
Figure 22 is success rate curve of the present embodiment on OTB100 test set under three-dimensional rotation scene, ordinate
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Figure 23 is that present embodiment removes the precision curve under sight scene, ordinate on OTB100 test set
Precision indicates precision, and abscissa Threshold indicates threshold value；
Figure 24 is that present embodiment removes the success rate curve under sight scene, ordinate on OTB100 test set
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold；
Figure 25 is precision curve of the present embodiment on OTB100 test set under dimensional variation scene, ordinate
Precision indicates precision, and abscissa Threshold indicates threshold value；
Figure 26 is success rate curve of the present embodiment on OTB100 test set under dimensional variation scene, ordinate
Successrate indicates success rate, and abscissa Overlap threshold indicates anti-eclipse threshold.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, those of ordinary skill in the art without creative labor it is obtained it is all its
His embodiment, shall fall within the protection scope of the present invention.
It should be noted that in the absence of conflict, the feature in embodiment and embodiment in the present invention can phase
Mutually combination.
The present invention will be further explained below with reference to the attached drawings and specific examples, but not as the limitation of the invention.
As shown in Figure 1, the characteristic detection method of the invention based on sparse coding, which comprises
S1, using extracting local feature region based on the FAST Corner Detection Algorithm of local threshold, and calculate characteristic point
Local gradient direction, using characteristic point as center sampled images block, as training sample；
S2, it training is grouped to dictionary element according to the Local gradient direction of characteristic point obtained complete dictionary；
S3, sparse spy is constructed then to image block piecemeal to test sample progress rarefaction representation using obtained dictionary
Sign realizes target following according to the detector of the sparse features.
In the S1 of present embodiment, the local feature region of former frame samplings is extracted, as training sample, uses present frame in S3
Sampling is used as test sample.
The core concept of the FAST Corner Detection Algorithm of present embodiment is the point of measuring point more to be checked He its peripheral region
Gray value, find out obvious brighter or darker point as characteristic point.
The S1 of present embodiment includes:
S11, image to be detected is switched to grayscale image, measuring point c to be checked is located at circle center, and surrounding point is located at circle void
On line, the gray value of measuring point to be checked is Ic, the gray value of surrounding point is Ic→p, the threshold value compared for gray value is T, such as
It is more than threshold value T that fruit, which has the difference of N number of gray value continuously put and measuring point to be checked, then measuring point to be checked is a characteristic point；Its
Middle threshold value T:
Wherein, C indicates the number of the point in 3 × 3 neighborhood of measuring point to be checked, IiIt is the gray scale put in 3 × 3 neighborhood of measuring point to be checked
Value, a indicate threshold factor, σ3×3Indicate the variance of gray value in neighborhood；Reaction
Be confidence level to threshold value, value range is [0,1]；
FAST detection algorithm uses a fixed threshold value.However, it may be desirable to characteristic point be regional area gray value become
Change more violent point.In general, the gray value of the pixel of a pixel and its eight neighborhood is close.If to be checked
Point difference in the gray value and neighborhood of measuring point is obvious, then the point may be affected by noise.Then consider the threshold of the increase point
Value.Conversely, the threshold value of the point should be relatively small.
According to the analysis of front, if the local variance of a point is bigger, threshold value also should correspondingly increase.In addition,
The factorWhat is reacted is the confidence level to threshold value, indicates central point and its neighborhood
The distribution situation of the difference of the gray value of interior point.Value range be [0,1].
If only one difference comparsion is big, value 0；If all differences are all, value 1.If in neighborhood only
The difference comparsion of some point is big, then central point is possible for characteristic point, accordingly decreases the factor to reduce threshold value.
S12, the gradient direction for calculating characteristic point:
△Ix=Ix+1-Ix-1
△Iy=Iy+1-Iy-1
Wherein, △ IxIndicate lateral gray value difference, △ IyIndicate longitudinal difference, ori indicates gradient direction；
Gradient orientation histogram at S13, statistical nature point, by gradient direction discretization to 0 °, 45 °, 90 °, 135 ° 4
On section, the image block of 4 Local gradient directions of characteristic point is obtained.
The S2 of present embodiment includes:
4 sub- dictionaries, trained 4 sub- dictionary groups is respectively trained using the image block of 4 different Local gradient directions
At excessively complete dictionary D；
Dictionary learning problem representation are as follows:
Wherein, Y=[y1,…,yk] it is the matrix that k training sample forms, each column vector ykIt is arranged by image block according to column
The mode conversion dimension of sequence obtains, X=[x1,…,xk] it is sparse coefficient matrix,Be to be solved it is excessively complete
Standby dictionary；K refers to degree of rarefication, i.e., each column vector x in sparse coefficient matrixiThe number of middle nonzero element, | | xi||0Indicate 0 model
Number；
Existing K-SVD algorithm calculates optimal dictionary and needs a large amount of matrix operation, is based on so present embodiment uses
The gradient of momentum declines to update the element in dictionary D:
Wherein, diIndicate the i-th column vector of dictionary D,Indicate sparse coefficient matrix the i-th row nonzero element composition row to
Amount,It is that dictionary element d can be used in matrix YiThe matrix for the training sample composition being indicated, it is non-with the i-th row of sparse matrix
Neutral element is corresponding,Indicate F norm；
Increment of sample vtUpdate method iteratively solves optimal dictionary element d, and then obtains complete dictionary D:
Wherein, v0It is the initial value of incremental update, vtIt is that intermediate variable is used for diIt is updated.μ is super on [0,1]
Parameter, step-length α are the learning rates in gradient decline.In order to determine optimal step size, function of a single variable of the minimum about α:
Derivative are as follows:
Wherein,It is second dervative.It is zero by derivative, available optimal step size:
The S3 of present embodiment includes:
S31, image block is acquired at each pixel of test sample image；
S32, rarefaction representation is carried out to image block by orthogonal matching pursuit OMP algorithm using obtained complete dictionary, obtained
Obtain the sparse coefficient of pixel in image block；
S33, to image block piecemeal:
As shown in Fig. 2, image block to be divided into 4 × 4 small cube, 8 × 8 image block each in this way is by 44 × 4 small sides
Block composition, 4 small cubes regard the small cube of 8 × 8 image block upper lefts, lower-left, upper right, bottom right as respectively；
S34, according to the position of each pixel, the sparse coefficient of the pixel projected to using bilinear interpolation around
4 small cubes on, the corresponding sparse coefficient of 4 small cubes is obtained according to the weight that projects on 4 small cubes；
4 small cubes around described include that small cube where the pixel is outer and adjacent with the small cube three images
The nearest small cube of distance in block, 8 × 8 image blocks of 4 small cubes composition；As shown in Figure 2: pixel is located at small cube
In D, then four small cubes of pixel projection are as follows: small cube D, the right side small cube adjacent with small cube D, downside square
And 8 × 8 image blocks that side small cube forms obliquely downward；
Project to the weight on 4 small cubes:
wcell1=dx2 × dy2
wcell2=dx1 × dy2
wcell3=dx2 × dy1
wcell4=dx1 × dy1
Wherein, dx1 and dx2 is left and right lateral distance of the pixel from four nearest small cube centers, and dy1 and dy2 are
Up and down fore-and-aft distance of the pixel from four nearest small cube centers；
As shown in Fig. 2, the image-region given for one, with a rectangular window slip scan, in each pixel
Place's acquisition image block.Then according to obtained complete dictionary D, sparse coding is carried out to these image blocks using OMP algorithm.It completes
After sparse coding, the feature histogram of square is constructed according to the rarefaction representation of pixel each in image-region.Original image
Each point in region corresponds to the sparse coefficient vector of n dimension.This sparse vector, which is considered as one, n bin
Histogram, then in the vector nonzero element carry out bilinear interpolation, carry out data smoothing.Each small cube is divided into
A, tetra- parts B, C, D.When carrying out interpolation, according to position of the pixel in block, by its rarefaction representation vector according to certain
Weight project on nearest block, the feature histogram for statistics block.In to scheme for the pixel of black, it is located at D
In, tetra- from cell1~cell4 cell blocks are nearest, so needing to carry out interpolation to this four cell blocks.Dx1 and dx2 is pixel
Lateral distance o'clock from four cell block centers, and dy1 and dy2 is longitudinal distance.These distances are both relative to cell block
Size p is normalized, therefore value interval is [0,1], dx1+dx2=1, dy1+dy2=1.
S35, the feature vector that each small cube is constructed according to the corresponding sparse coefficient of 4 small cubes that S34 is obtained, the spy
Levying vector is three-dimensional eigenmatrix, and longitudinal dimension of the matrix normalizes to obtain final sparse features；
S36, target following is realized according to the detector of the sparse features.
Present embodiment proposes a kind of sparse features detection method based on sparse coding, to improve the Shandong of target following
Stick.Its basic thought is carried out dictionary learning to the training sample of former frame samplings and obtains dictionary, then to the survey of present frame
Sample this progress sparse features expression.It carries out the adaptive FAST characteristic point of local threshold to image block first to detect, then with spy
Acquisition image block obtains training sample centered on sign point.By calculating the Local gradient direction of image block, to the element of dictionary into
Row station work, to describe the local direction information of different images block.Finally, image block is divided into small cube, by sparse spy
Sign projects in these small cubes, reconstructs the feature vector of block.Tracking process estimates the position of target using correlation filter
It sets.The experimental results showed that especially under rotation scene, the algorithm of proposition is in terms of tracking accuracy on OTB100 test set
Better than current main-stream track algorithm.
If Fig. 3 and Fig. 4 are respectively precision curve and success rate curve of the present embodiment on OTB100 test set.
OTB100 test set includes 100 videos, these videos can substantially be divided into the challenging scene of 11 classes: background is obscured
(background clutters, BC), deformation (deformation, DEF) quickly move (fast motion, FM), illumination
Change (illumination variation, IV), plane internal rotation (in-plane rotation, IPR), low resolution
(low resolution, LR), motion blur (motion blur, MB), blocks (occlusion, OCC), three-dimensional rotation
(out-of-plane rotation, OPR) is removed sight (out-of-view, OV), dimensional variation (scale
Variation, SV).As can be seen that the track algorithm of present embodiment is better than other assessment algorithms in performance.Compared to core
Correlation filtering KCF algorithm, since the algorithm of present embodiment and the difference of KCF are that feature detection scheme is inconsistent, institute
Can determine that the feature detection algorithm based on sparse coding is conducive to improve the accuracy of detection, and then improve tracking performance.
The accuracy of GOTURN algorithm based on deep learning will be lower than other algorithms.This is because GOTURN passes through outside learning objective
The connection with motion conditions is seen to estimate the position of target, and position possible for the image object of a M × N has (MN)2
In, it is difficult to the physical location of directly accurate estimation target.In figure, the method that Ours indicates present embodiment, SRDCF representation space
Regularized discriminant correlation filter, CN indicate the detection algorithm based on color characteristic, CMT (Clustering of Static-
Adaptive Correspondences for Deformable Object Tracking) it is a kind of target tracking algorism；
If Fig. 5 to Figure 26 is precision and success rate curve of the present embodiment on OTB100 test set under different scenes.
Under various scenes, the track algorithm that present embodiment proposes all is first two, and performance is considerably beyond KCF algorithm.This card
The sparse features detector proposed is illustrated to be conducive to improve detection performance and track robustness.For IPR, LR, OCC, OPR and SV
Scene, the accuracy highest of the algorithm of proposition.Under IPR and OPR scene, target and background constantly rotates.Due to this embodiment party
The dictionary for the sparse features detector that formula obtains is to update dictionary element according to the local direction of image block, is had stronger
Rotational invariance, so having very high robustness in the case where rotating scene.It corresponds, other algorithms are for target direction
Change more sensitive.Main cause be HOG algorithm statistics be target Local gradient direction histogram, target direction information
Variation will have a direct impact on the distribution situation of histogram.The method that present embodiment proposes also has very high under LR and SV scene
Accuracy has very local detail this is because the sparse features learnt are able to reflect the partial structurtes information of target
Strong expression ability.This illustrates that present embodiment characteristic detection method ratio HOG has better character representation ability, can be continuous
The new feature of learning objective, adapts to the variation of target appearance, is conducive to improve detection performance.And HOG feature is due to being to set by hand
Meter, target can only be described according to fixed mode.So compared to existing HOG, the sparse features detection algorithm of proposition
Be conducive to improve the accuracy and robustness of track algorithm.
Although describing the present invention herein with reference to specific embodiment, it should be understood that, these realities
Apply the example that example is only principles and applications.It should therefore be understood that can be carried out to exemplary embodiment
Many modifications, and can be designed that other arrangements, without departing from spirit of the invention as defined in the appended claims
And range.It should be understood that different appurtenances can be combined by being different from mode described in original claim
Benefit requires and feature described herein.It will also be appreciated that the feature in conjunction with described in separate embodiments can be used
In other described embodiments.
Claims (5)
1. a kind of characteristic detection method based on sparse coding, which is characterized in that the described method includes:
S1, using extracting local feature region based on the FAST Corner Detection Algorithm of local threshold, and calculate the part of characteristic point
Gradient direction, using characteristic point as center sampled images block, as training sample；
S2, it training is grouped to dictionary element according to the Local gradient direction of characteristic point obtained complete dictionary；
S3, rarefaction representation is carried out to test sample using obtained dictionary, then to image block piecemeal, constructs sparse features, benefit
Target following is realized with the detector of the sparse features.
2. the characteristic detection method according to claim 1 based on sparse coding, which is characterized in that in S1, extract former
The local feature region of frame sampling uses current frame sampling as test sample as training sample in S3.
3. the characteristic detection method according to claim 1 or 2 based on sparse coding, which is characterized in that the S1 includes:
S11, image to be detected is switched to grayscale image, measuring point c to be checked is located at circle center, and surrounding point is located at circled hash
On, the gray value of measuring point to be checked is Ic, the gray value of surrounding point is Ic→p, the threshold value compared for gray value is T, if
The difference for having N number of gray value continuously put and measuring point to be checked is more than threshold value T, then measuring point to be checked is a characteristic point；Wherein
Threshold value T:
Wherein, C indicates the number of the point in 3 × 3 neighborhood of measuring point to be checked, IiIt is the gray value put in 3 × 3 neighborhood of measuring point to be checked, a
Indicate threshold factor, σ 3×3Indicate the variance of gray value in neighborhood；Reaction be
To the confidence level of threshold value, value range is [0,1]；
S12, the gradient direction for calculating characteristic point:
△Ix=Ix+1-Ix-1
△Iy=Iy+1-Iy-1
Wherein, △ IxIndicate lateral gray value difference, △ IyIndicate longitudinal difference, ori indicates gradient direction；
Gradient orientation histogram at S13, statistical nature point, by gradient direction discretization to 0 °, 45 °, 90 °, 135 ° of 4 sections
On, obtain the image block of 4 Local gradient directions of characteristic point.
4. the characteristic detection method according to claim 3 based on sparse coding, which is characterized in that
The S2 includes:
4 sub- dictionaries are respectively trained using the image block of 4 different Local gradient directions, trained 4 sub- dictionaries formed
Complete dictionary D；
Dictionary learning problem representation are as follows:
Wherein, Y=[y1,…,yk] it is the matrix that k training sample forms, each column vector ykBy image block according to column sequence
Mode conversion dimension obtains, X=[x1,…,xk] it is sparse coefficient matrix,Be to be solved it is excessively complete
Dictionary；K refers to degree of rarefication, i.e., each column vector x in sparse coefficient matrixiThe number of middle nonzero element, | | xi||0Indicate 0 model
Number；
The element in dictionary D is updated using based on the decline of the gradient of momentum:
Wherein, diIndicate the i-th column vector of dictionary D,Indicate the row vector of sparse coefficient matrix the i-th row nonzero element composition,
It is that dictionary element d can be used in matrix YiThe matrix for the training sample composition being indicated, the i-th row non-zero entry with sparse matrix
Element is corresponding,Indicate F norm；
Increment of sample vtUpdate method iteratively solves optimal dictionary element d, and then obtains complete dictionary D:
Wherein, v0It is the initial value of incremental update, vtIt is that intermediate variable is used for diIt is updated；μ is the hyper parameter on [0,1],
5. the characteristic detection method according to claim 4 based on sparse coding, which is characterized in that the S3 includes:
S31, image block is acquired at each pixel of test sample image；
S32, rarefaction representation is carried out to image block by orthogonal matching pursuit OMP algorithm using obtained complete dictionary, is schemed
As the sparse coefficient of pixel in block；
S33, to image block piecemeal:
Image block is divided into 4 × 4 small cube, 8 × 8 image block each in this way is made of 44 × 4 small cubes, 4 small sides
Block regards the small cube of 8 × 8 image block upper lefts, lower-left, upper right, bottom right as respectively；
S34, according to the position of each pixel, the sparse coefficient of the pixel is projected to the 4 of surrounding using bilinear interpolation
On a small cube, the corresponding sparse coefficient of 4 small cubes is obtained according to the weight projected on 4 small cubes；
4 small cubes around described include outside the small cube of the pixel place and in three image blocks adjacent with the small cube
Apart from nearest small cube, 8 × 8 image blocks of 4 small cubes composition；
Project to the weight on 4 small cubes:
wcell1=dx2 × dy2
wcell2=dx1 × dy2
wcell3=dx2 × dy1
wcell4=dx1 × dy1
Wherein, dx1 and dx2 is left and right lateral distance of the pixel from four nearest small cube centers, and dy1 and dy2 are pixel
Up and down fore-and-aft distance of the point from four nearest small cube centers；
S35, construct the feature vector of each small cube according to the corresponding sparse coefficient of 4 small cubes that S34 is obtained, this feature to
Amount is three-dimensional eigenmatrix, and longitudinal dimension of the matrix normalizes to obtain final sparse features；
S36, target following is realized according to the detector of the sparse features.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811021166.6A CN109241981B (en) | 2018-09-03 | 2018-09-03 | Feature detection method based on sparse coding |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811021166.6A CN109241981B (en) | 2018-09-03 | 2018-09-03 | Feature detection method based on sparse coding |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109241981A true CN109241981A (en) | 2019-01-18 |
| CN109241981B CN109241981B (en) | 2022-07-12 |
Family
ID=65060353
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811021166.6A Active CN109241981B (en) | 2018-09-03 | 2018-09-03 | Feature detection method based on sparse coding |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109241981B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110648368A (en) * | 2019-08-30 | 2020-01-03 | 广东奥普特科技股份有限公司 | Calibration board corner point discrimination method based on edge features |
| CN112699829A (en) * | 2021-01-05 | 2021-04-23 | 山东交通学院 | Vehicle weight identification method and system based on depth feature and sparse measurement projection |
| CN113256685A (en) * | 2021-06-25 | 2021-08-13 | 南昌工程学院 | Target tracking method and system based on convolutional neural network dictionary pair learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103679216A (en) * | 2014-01-06 | 2014-03-26 | 重庆大学 | Method for detecting infrared small/weak targets based on space-time sparse representation |
| CN104616291A (en) * | 2015-01-15 | 2015-05-13 | 东华大学 | Sparse coding-based fabric appearance flatness evaluation method |
| US20150227819A1 (en) * | 2014-02-07 | 2015-08-13 | Canon Kabushiki Kaisha | Image processing method and apparatus using trained dictionary |
| US20160253782A1 (en) * | 2013-11-01 | 2016-09-01 | Eizo Corporation | Image quality enhancing apparatus, image display apparatus, image quality enhancing method, and computer readable storage medium |
| CN106529526A (en) * | 2016-07-06 | 2017-03-22 | 安徽大学 | Object tracking algorithm based on combination between sparse expression and prior probability |
-
2018
- 2018-09-03 CN CN201811021166.6A patent/CN109241981B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160253782A1 (en) * | 2013-11-01 | 2016-09-01 | Eizo Corporation | Image quality enhancing apparatus, image display apparatus, image quality enhancing method, and computer readable storage medium |
| CN103679216A (en) * | 2014-01-06 | 2014-03-26 | 重庆大学 | Method for detecting infrared small/weak targets based on space-time sparse representation |
| US20150227819A1 (en) * | 2014-02-07 | 2015-08-13 | Canon Kabushiki Kaisha | Image processing method and apparatus using trained dictionary |
| CN104616291A (en) * | 2015-01-15 | 2015-05-13 | 东华大学 | Sparse coding-based fabric appearance flatness evaluation method |
| CN106529526A (en) * | 2016-07-06 | 2017-03-22 | 安徽大学 | Object tracking algorithm based on combination between sparse expression and prior probability |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110648368A (en) * | 2019-08-30 | 2020-01-03 | 广东奥普特科技股份有限公司 | Calibration board corner point discrimination method based on edge features |
| CN110648368B (en) * | 2019-08-30 | 2022-05-17 | 广东奥普特科技股份有限公司 | Calibration board corner point discrimination method based on edge features |
| CN112699829A (en) * | 2021-01-05 | 2021-04-23 | 山东交通学院 | Vehicle weight identification method and system based on depth feature and sparse measurement projection |
| CN112699829B (en) * | 2021-01-05 | 2022-08-30 | 山东交通学院 | Vehicle weight identification method and system based on depth feature and sparse measurement projection |
| CN113256685A (en) * | 2021-06-25 | 2021-08-13 | 南昌工程学院 | Target tracking method and system based on convolutional neural network dictionary pair learning |
| CN113256685B (en) * | 2021-06-25 | 2021-09-24 | 南昌工程学院 | Target tracking method and system based on convolutional neural network dictionary pair learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109241981B (en) | 2022-07-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wu et al. | ORSIm detector: A novel object detection framework in optical remote sensing imagery using spatial-frequency channel features | |
| Chen et al. | Research on recognition method of electrical components based on YOLO V3 | |
| US10607362B2 (en) | Remote determination of containers in geographical region | |
| CN110543837B (en) | Visible light airport airplane detection method based on potential target point | |
| Tareef et al. | Optimizing the cervix cytological examination based on deep learning and dynamic shape modeling | |
| CN107657279B (en) | Remote sensing target detection method based on small amount of samples | |
| CN109684925B (en) | Depth image-based human face living body detection method and device | |
| Jia et al. | Visual tracking via adaptive structural local sparse appearance model | |
| CN107133627A (en) | Infrared light spot center point extracting method and device | |
| CN106485651B (en) | The image matching method of fast robust Scale invariant | |
| CN104299260A (en) | Contact network three-dimensional reconstruction method based on SIFT and LBP point cloud registration | |
| Lei et al. | End-to-end change detection using a symmetric fully convolutional network for landslide mapping | |
| CN111275686B (en) | Method and device for generating medical image data for artificial neural network training | |
| CN108573231B (en) | Human body behavior identification method of depth motion map generated based on motion history point cloud | |
| CN110334762A (en) | A kind of feature matching method combining ORB and SIFT based on quaternary tree | |
| Woźniak et al. | Graphic object feature extraction system based on cuckoo search algorithm | |
| CN109241981A (en) | A kind of characteristic detection method based on sparse coding | |
| CN103985143A (en) | Discriminative online target tracking method based on videos in dictionary learning | |
| CN110245587B (en) | Optical remote sensing image target detection method based on Bayesian transfer learning | |
| CN105488541A (en) | Natural feature point identification method based on machine learning in augmented reality system | |
| Bu et al. | Mask-CDNet: A mask based pixel change detection network | |
| EP3553700A2 (en) | Remote determination of containers in geographical region | |
| Zhong et al. | Robust tracking via patch-based appearance model and local background estimation | |
| An et al. | Transitive transfer learning-based anchor free rotatable detector for SAR target detection with few samples | |
| CN115984186A (en) | Fine product image anomaly detection method based on multi-resolution knowledge extraction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |