CN109584273A - A kind of moving target detecting method based on adaptive convergence parameter - Google Patents
A kind of moving target detecting method based on adaptive convergence parameter Download PDFInfo
- Publication number
- CN109584273A CN109584273A CN201811389171.2A CN201811389171A CN109584273A CN 109584273 A CN109584273 A CN 109584273A CN 201811389171 A CN201811389171 A CN 201811389171A CN 109584273 A CN109584273 A CN 109584273A
- Authority
- CN
- China
- Prior art keywords
- matrix
- infrared image
- rank
- frame
- sparse
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10048—Infrared image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Abstract
The invention belongs to technical field of image processing, disclose a kind of moving target detecting method based on adaptive convergence parameter.This method comprises the following steps: first with the infrared image of thermal infrared imager continuous collecting area to be tested, then by adaptively being adjusted to preset initial convergence parameter, obtain adaptive convergence parameter, then continuous infrared image is trained using Robust Principal Component Analysis method, obtain the corresponding sparse image of each frame infrared image, and then the corresponding sparse image of each frame infrared image of Sequential output, complete moving object detection.Method provided by the invention can be before being trained continuous infrared image using Robust Principal Component Analysis method, preset initial convergence parameter is adaptively adjusted, and then when being trained to continuous infrared image to obtain the corresponding sparse matrix of each frame infrared image, the number of iterations of training process can be reduced, operand is reduced, operation efficiency is improved.
Description
Technical field
The present invention relates to technical field of image processing more particularly to a kind of moving target inspections based on adaptive convergence parameter
Survey method.
Background technique
Infrared electronic technology based on infrared sensor has the ability to work of round-the-clock, thus is widely used in military affairs
And civil field.Meanwhile Detection for Moving Target is always the important topic of image procossing and computer vision research, equally
It has a wide range of applications in many fields.Wherein, the temperature information for the moving target that infrared sensor extracts be moving target with
One of the important feature that static scene is distinguished, therefore, the thermal infrared images formed to the temperature information of target carry out relevant treatment,
Moving target and static scene can be distinguished, realize the detection of moving target.
However under complex electromagnetic environment, the background of moving target is also more complex, moving object detection and tracking face as
Lower challenge and problem: (1) when moving target scale smaller, such as sub-mini unmanned rotorcraft, far field players, it is infrared at
As to occupy infrared image pixel also less for moving target in rear data;(2) moving target and environment temperature contrast are small, cause red
Aimless factors interference is big in data after outer imaging.For the above problem, scholar both domestic and external has carried out a series of research, mentions
A kind of Robust Principal Component Analysis method is gone out, this method can be by the corresponding input number of continuous infrared image from Video Quality Metric
It is low-rank matrix and sparse matrix according to matrix decomposition, sparse matrix therein just represents the matrix of moving target, but this method
In decomposable process, the number of iterations is more, causes operand big, operation efficiency is low.
Summary of the invention
In view of this, the present invention provides a kind of movement mesh detection method based on adaptive convergence parameter, can use
Before Robust Principal Component Analysis method is trained continuous infrared image, for different area to be tested, to preset initial
Convergence parameter is adaptively adjusted, so can reduction the number of iterations in the training process, reduce operand, improve operation
Efficiency.
In order to achieve the above objectives, the embodiment of the present invention adopts the following technical scheme that
A kind of movement mesh detection method based on adaptive convergence parameter is provided, comprising the following steps:
Step 1, the infrared image of thermal infrared imager continuous collecting area to be tested is utilized.
Step 2, the initial 4 frame infrared image for obtaining thermal infrared imager acquisition, using initial 4 frame infrared image to preset
Initial convergence parameter ε is adaptively adjusted, and adaptive convergence parameter ε ' is obtained.
Step 3, for thermal infrared imager acquisition every continuous infrared image of n frame, and using adaptive convergence parameter ε ' as
Convergence parameter is trained every continuous infrared image of n frame using robust principal component analysis method, it is continuously infrared to obtain every n frame
The corresponding sparse matrix of each frame infrared image in image;Wherein, n is integer, n >=4.
Step 4, according to the corresponding sparse matrix of each frame infrared image, the corresponding sparse graph of each frame infrared image is obtained
Picture, and then the corresponding sparse image of each frame infrared image of Sequential output, wherein gray value is not zero in each sparse image
Region, that is, moving target region that pixel is constituted.
Moving target detecting method provided by the invention based on adaptive convergence parameter, is adopted first with thermal infrared imager
The infrared image for collecting area to be tested, then adaptively adjusts preset initial convergence parameter, is adaptively restrained
Parameter;Then continuous infrared image is trained using Robust Principal Component Analysis method, is obtained each in continuous infrared image
The corresponding sparse matrix of frame infrared image, and then the corresponding sparse image of each frame infrared image is obtained, wherein each sparse graph
The region i.e. moving target region that the pixel that gray value is not zero as in is constituted, the sequence according to acquisition infrared image are defeated
The corresponding sparse image of each frame infrared image out, wherein the area that the pixel that gray value is not zero in each sparse image is constituted
Domain, that is, moving target region, by observation you can learn that the movement tendency of moving target.
Method provided by the invention is directed to different area to be tested, using 4 initial frame infrared images to preset first
Beginning convergence parameter is adaptively adjusted, and adaptive convergence parameter is obtained, and using adaptive convergence parameter as convergence parameter, is used
Robust Principal Component Analysis is trained continuous infrared image, and then obtains the corresponding sparse matrix of each frame infrared image,
So, since with adaptive adjusted convergence parameter, the number of iterations of training process can be reduced in the training process,
Operand is reduced, operation efficiency is improved.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is a kind of process of the moving target detecting method based on adaptive convergence parameter provided in an embodiment of the present invention
Schematic diagram;
Fig. 2 is the grayscale image of the infrared image of area to be tested;
Fig. 3 is to handle the 10th frame, the 20th frame, the 30th frame and the 40th infrared figure of frame using method provided in an embodiment of the present invention
As obtained testing result figure, wherein Fig. 3 (a) is the corresponding sparse image of the 10th frame infrared image, and Fig. 3 (b) is that the 20th frame is red
The corresponding sparse image of outer image, Fig. 3 (c) are the corresponding sparse image of the 30th frame infrared image, and Fig. 3 (d) is that the 40th frame is infrared
The corresponding sparse image of image;
Fig. 4 is to handle the sparse image and its office that a certain frame infrared image obtains using method provided in an embodiment of the present invention
Portion's enlarged drawing;
Fig. 5 is the root-mean-square error and operation time change curve of Robust Principal Component Analysis method under different convergence parameters.
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, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
Fig. 1 is a kind of process of the moving target detecting method based on adaptive convergence parameter provided in an embodiment of the present invention
Schematic diagram.
Referring to Fig. 1, the moving target detecting method provided in an embodiment of the present invention based on adaptive convergence parameter include with
Lower step:
Step 1, the infrared image of thermal infrared imager continuous collecting area to be tested is utilized.
Step 2, the initial 4 frame infrared image for obtaining thermal infrared imager acquisition, using initial 4 frame infrared image to preset
Initial convergence parameter ε is adaptively adjusted, and adaptive convergence parameter ε ' is obtained.
Specifically, step 2 the following steps are included:
(2.1) it initializes: obtaining the corresponding input data matrix D of initial 4 frame infrared image;I is enabled to indicate interative computation
Number;EiIndicate the sparse matrix in i-th iteration operation, E0For the matrix and E of the rank of m × 40In element all 0;AiIt indicates
Low-rank matrix in i-th iteration operation, A0For the matrix and A of the rank of m × 40In element all 0;λ is indicated to sparse matrix
The weight of imparting,YiIndicate the Lagrange multiplier matrix in the interative computation of i-th, Y0=D/max (| | D |
|2,λ-1||D||∞);μiIndicate Y in i-th iteration operationiPenalty factor, μ0=1.0/ | | D | |2;ρ is μkRegulatory factor,
0.001≤ρ≤5;ε is initial convergence parameter, ε=1 × 10-7;The adjustment number of initial convergence parameter, I=0 are indicated with I.
Wherein, the corresponding input data matrix D of initial 4 frame infrared image is the rank matrix of m × 4,H is image
Height, W is the width of image, and H and W are even number, | | D | |2Two norms of input matrix D are sought in expression, | | D | |∞Expression asks defeated
Enter the Infinite Norm of matrix D, max () indicates to take the maximum value of element in matrix in bracket.
(2.2) i=1 is enabled.
(2.3) the sparse matrix E in i-th iteration operation is calculatedi:
Wherein, EiFor the rank matrix of m × 4, B is the rank matrix of m × 4, each element is λ/μ in matrix Bi-1。
(2.4) the pilot process matrix Temp in i-th iteration operation is calculatedi:
Tempi=D-Ei+μi-1Yi-1;
Wherein, TempiFor the rank matrix of m × 4.
(2.5) to pilot process matrix TempiCarry out singular value decomposition;
(Ui,∑i,Vi)=svd (Tempi),
Wherein UiFor m × m rank matrix, indicate to pilot process matrix TempiCarry out the left You Te obtained after singular value decomposition
Levy vector matrix;ViFor 4 × 4 rank matrixes, ViIt indicates to pilot process matrix TempiCarry out the right tenth of the twelve Earthly Branches obtained after singular value decomposition
Feature matrix;∑iFor the rank matrix of m × 4, ∑iIt indicates to pilot process matrix TempiIt is obtained after progress singular value decomposition
Singular value matrix;Svd indicates to carry out singular value decomposition to matrix.
(2.6) the sparse matrix A in i-th iteration operation is calculatedi:
Αi=Ui(∑i-μi)Vi T,
Wherein, AiFor the rank matrix of m × 4, the transposition of subscript T representing matrix, μiFor the rank matrix of m × 4, μiEach element
It is 1/ μ in i-th iterationi-1。
(2.7) the Lagrange multiplier matrix Y in i-th iteration operation is calculatedi:
Yi=Yi-1+μi-1(D-Ei-Ai),
Wherein, YiFor the rank matrix of m × 4.
The corresponding sound disintegrate-quality relative error ξ of i-th iteration operation is calculatedi=| | D-Ei-Ai||1/||D||1,
Judge ξiWhether ε is greater than: if ξiGreater than ε, then μ is calculatedi=min (μi-1×ρ,μi-1×10-7), it enables i add 1, executes step
(2.3);If ξiLess than or equal to ε, judge whether I is equal to 0, if I is equal to 0, thens follow the steps (2.8);If I is not equal to 0, hold
Row step (2.9);Wherein, min () expression is minimized.
(2.8) E is enabledS=Ei, calculate the initial value of root-mean-square errorIt enables I add 1, and calculates
The ε of ε '=10 is obtained, ε is enabled to take ε ', is executed step (2.2).
Wherein, | | | |1The 1- norm of representing matrix, sum () indicate to sum to all elements of column vector in bracket, Ei
(:, 1) it indicates to take matrix E1First row, Es(:, 1) it indicates to take matrix ESFirst row, length () expression seek square in bracket
The element number of battle array, ones (length (Ei), 1) indicate element be 1 and order be (length (Ei(:, 1)), 1) × 1 square
Battle array.
(2.9) the sparse matrix E in i-th this interative computation is calculatediCorresponding root-mean-square error:Calculate |
RMSEi-S|;If | RMSEi- S | less than 1, then the ε of ε '=10, enables ε take ε ', and I is enabled to add 1, executes step (2.2);If | RMSEi-
S | it is more than or equal to 1, then the ε of ε '=0.1, obtains adaptive convergence parameter ε '.
It should be noted that the sparse matrix that root-mean-square error RMSE is calculated in the case where referring to convergence parameter difference
With the root-mean-square error between effective sparse matrix, the present invention with the adjustment number of initial convergence parameter be 0 when be calculated
Sparse matrix be effective sparse matrix.
Preferably, the specific generating mode of input data matrix D are as follows:
Infrared image continuous for l frame obtains the corresponding data matrix D of each frame infrared image1,D2,…,Dl;It extracts every
The even number line of the corresponding data matrix of one frame infrared image and the data of even column obtain adopting under each frame infrared image is corresponding
Sample data matrix D1',D2',…,Dl';The corresponding down-sampled data matrix of each frame infrared image is stretched as vector according to column,
Obtain the corresponding column vector of the corresponding down-sampled data matrix of each frame infrared imageAnd then obtain input data
Matrix
Wherein, l >=4, the element in data matrix are the pixel value of each frame infrared image corresponding position, data matrix
D1,D2,…,DlOrder be H × W;Down-sampled data matrix order is Order be m × 1, it is defeated
The order for entering data matrix D is m × l.
Step 3, for thermal infrared imager acquisition every continuous infrared image of n frame, and using adaptive convergence parameter ε ' as
Convergence parameter is trained every continuous infrared image of n frame using robust principal component analysis method, it is continuously infrared to obtain every n frame
The corresponding sparse matrix of each frame infrared image in image;Wherein, n is integer, n >=4.
Specifically, step 3 the following steps are included:
(3.1) it initializes: obtaining the corresponding input data matrix D of every continuous infrared image of n frame;K indicates the number of iterations, Ek
Indicate the sparse matrix in kth time interative computation, E0For the matrix and E of m × n rank0In element all 0;AkIndicate kth time
Low-rank matrix in interative computation, A0For the matrix and A of m × n rank0In element all 0;λ is indicated in intra-frame trunk algorithm
To sparse matrix EkThe weight of imparting,YkIndicate the Lagrange multiplier matrix in the interative computation of kth time, Y0
=D/max (| | D | |2,λ-1||D||∞);μkIndicate Y in kth time interative computationkPenalty factor, μ0=1.0/ | | D | |2;ρ is μk
Regulatory factor, 0.001≤ρ≤5.
(3.2) k=1 is enabled.
(3.3) the sparse matrix E in kth time interative computation is calculatedk:
Wherein, EkFor m × n rank matrix, B is m × n rank matrix, each element is λ/μ in matrix Bk-1。
(3.4) to pilot process matrix TempkCarry out singular value decomposition:
Tempk=D-Ek+μk-1Yk-1;
Wherein, TempkFor m × n rank matrix.
(3.5) to pilot process matrix TempkCarry out different value decomposition
(Uk,∑k,Vk)=svd (Tempk),
Wherein, wherein UkFor m × m rank matrix, indicate to pilot process matrix TempkIt is obtained after progress singular value decomposition
Left tenth of the twelve Earthly Branches feature matrix;VkFor n × n rank matrix, VkIt indicates to pilot process matrix TempkIt is obtained after carrying out singular value decomposition
Right tenth of the twelve Earthly Branches feature matrix;∑kFor m × n rank matrix, ∑kIt indicates to pilot process matrix TempkAfter carrying out singular value decomposition
Obtained singular value matrix.
(3.6) the sparse matrix A in kth time interative computation is calculatedk:
Αk=Uk(∑i-μk)Vk T,
Wherein, AkFor m × n rank matrix, μkFor m × n rank matrix, μkEach element be in kth time iteration 1/
μk-1。
(3.7) the Lagrange multiplier matrix Y in the interative computation of kth time is calculatedk:
Yk=Yk-1+μk-1(D-Ek-Ak);
Wherein, YkFor m × n rank matrix.
(3.8) the corresponding sound disintegrate-quality relative error ξ of kth this interative computation is calculatedk=| | D-Ek-Ak||1/|
|D||1If ξk> ε ', then be calculated μk=min (μk-1×ρ,μk-1×10-7), it enables k add 1, then executes step (3.2);It is no
Then follow the steps (3.9).
(3.9) stop iteration, export sparse matrix Ek;
Wherein, EkFor m × n rank matrix.
(3.10) sparse matrix is enabledRestore column vectorObtain the corresponding sparse matrix E of each frame infrared image1,E2,…,En。
Wherein,Indicate p-th of column vector,Ep's
Order is h × w,
Step 4, according to the corresponding sparse matrix of each frame infrared image, the corresponding sparse graph of each frame infrared image is obtained
Picture, and then the corresponding sparse image of each frame infrared image of Sequential output, wherein gray value is not zero in each sparse image
Region, that is, moving target region that pixel is constituted.
It should be noted that the corresponding sparse graph of each frame infrared image of Sequential output seems to refer to, the continuous infrared figure of every n frame
The acquisition order for having oneself of picture presses sparse image wherein each frame infrared image has oneself corresponding sparse image
It is continuously exported according to the acquisition order of its infrared image for corresponding to frame, by observation moving target region you can learn that movement mesh
Target movement tendency.
4 initial frame infrared images pair are utilized for different area to be tested based on above scheme of the embodiment of the present invention
Preset initial convergence parameter is adaptively adjusted, and adaptive convergence parameter is obtained, using adaptive convergence parameter as convergence
Parameter is trained continuous infrared image using Robust Principal Component Analysis, and then it is corresponding to obtain each frame infrared image
Sparse matrix, so, since with adaptive adjusted convergence parameter, training process can be reduced in the training process
The number of iterations reduces operand, improves operation efficiency.
Further, the above-mentioned beneficial effect of the present invention is verified below by way of emulation experiment.
Experiment one:
It takes certain subsequent place of school teaching building as area to be tested, and uses the huge online thermal infrared imager of brother MAG-62
The infrared image of continuous collecting area to be tested obtains the infrared image of area to be tested, and carries out gray processing processing, obtains one
Series grayscale image similar to Figure 2.Wherein, the huge online thermal infrared imager of brother MAG-62 uses 17 μm, 640 × 480 non-systems
Cold focus planar detector, temperature resolution 40mk, each frame sign of the infrared image of area to be tested are 640 × 480, take n
=4.
The 40 frame infrared images for choosing thermal infrared imager acquisition, are handled using method provided by the invention, are obtained every
The corresponding sparse image of frame infrared image.Wherein, the 10th frame, the 20th frame, the 30th frame and the 40th frame infrared image are corresponding sparse
Image is respectively as shown in Fig. 3 (a), Fig. 3 (b), Fig. 3 (c) and Fig. 3 (d).
The sub-mini unmanned rotorcraft that observation Fig. 3 can be seen that the upper right corner has movement tendency to the left.Although microminiature
Rotor wing unmanned aerial vehicle scale is small, and calorific value is low, and occupied infrared image pixel is few, can detecte using method provided by the invention
This moving target out.The corresponding sparse image of a certain frame in 40 frame infrared image of arbitrary extracting, and in the sparse image
Moving target amplifies, as shown in Figure 4.From fig. 4, it can be seen that gyroplane is not only detected, and its shape contour with
Entity is almost the same, and objective contour is clear.
Experiment two:
The 40 frame infrared images that thermal infrared imager is acquired using Robust Principal Component Analysis method under different convergence parameters
In certain continuous infrared image of 4 frames handled, obtain the root-mean-square error of Robust Principal Component Analysis method under different convergence parameters
With operation time change curve, as shown in Figure 5.From fig. 5, it can be seen that when ε is initial convergence parameter, i.e. ε=1 × 10-7When,
Operation time is 0.923 second, and taking root-mean-square error at this time is initial root mean square error.Take initial root mean square error and root mean square
The absolute value of the difference of error is effective testing result less than 1, from fig. 5, it can be seen that guaranteeing the effective premise of testing result
Under, when ε is using the method for the present invention convergence parameter adjusted, i.e. ε=1 × 10-2When, operation time is 0.319 second.Obviously,
Operation time can be reduced under the premise of guaranteeing that testing result is effective using method provided in an embodiment of the present invention, improve operation
Efficiency.
Those of ordinary skill in the art will appreciate that: realize that all or part of the steps of above method embodiment can pass through
The relevant hardware of program instruction is completed, and program above-mentioned can be stored in a computer readable storage medium, the program
When being executed, step including the steps of the foregoing method embodiments is executed;And storage medium above-mentioned includes: ROM, RAM, magnetic disk or light
The various media that can store program code such as disk.
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
Those familiar with the art in the technical scope disclosed by the present invention, can easily think of the change or the replacement, and should all contain
Lid is within protection scope of the present invention.Therefore, protection scope of the present invention should be based on the protection scope of the described claims.
Claims (4)
1. a kind of moving target detecting method based on adaptive convergence parameter, which comprises the following steps:
Step 1, the infrared image of thermal infrared imager continuous collecting area to be tested is utilized;
Step 2, the initial 4 frame infrared image for obtaining the thermal infrared imager acquisition, using the initial 4 frame infrared image to pre-
If initial convergence parameter ε adaptively adjusted, obtain adaptive convergence parameter ε ';
Step 3, for every continuous infrared image of n frame of thermal infrared imager acquisition, and with the adaptive convergence parameter ε '
As convergence parameter, every continuous infrared image of n frame is trained using robust principal component analysis method, is obtained described every
The corresponding sparse matrix of each frame infrared image in the continuous infrared image of n frame;Wherein, n is integer, n >=4;
Step 4, according to the corresponding sparse matrix of each frame infrared image, the corresponding sparse image of each frame infrared image is obtained,
And then the corresponding sparse image of each frame infrared image of Sequential output, wherein the pixel that gray value is not zero in each sparse image
Region, that is, moving target region that point is constituted.
2. the method according to claim 1, wherein step 2 the following steps are included:
(2.1) it initializes: obtaining the corresponding input data matrix D of the initial 4 frame infrared image;I is enabled to indicate interative computation
Number;EiIndicate the sparse matrix in i-th iteration operation, E0For the matrix and E of the rank of m × 40In element all 0;AiIt indicates
Low-rank matrix in i-th iteration operation, A0For the matrix and A of the rank of m × 40In element all 0;λ is indicated to described sparse
The weight that matrix assigns,YiIndicate the Lagrange multiplier matrix in the interative computation of i-th, Y0=D/max (|
|D||2,λ-1||D||∞);μiIndicate Y in i-th iteration operationiPenalty factor, μ0=1.0/ | | D | |2;ρ is μkAdjusting because
Son, 0.001≤ρ≤5;ε is initial convergence parameter, ε=1 × 10-7;The adjustment number of initial convergence parameter, I=0 are indicated with I;
Wherein, the corresponding input data matrix D of the initial 4 frame infrared image is the rank matrix of m × 4,H is image
Height, W is the width of image, and H and W are even number, | | D | |2Two norms of the input matrix D are sought in expression, | | D | |∞It indicates
The Infinite Norm of the input matrix D is sought, max () indicates to take the maximum value of element in matrix in bracket;
(2.2) i=1 is enabled;
(2.3) the sparse matrix E in i-th iteration operation is calculatedi:
Wherein, EiFor the rank matrix of m × 4, B is the rank matrix of m × 4, each element is λ/μ in matrix Bi-1;
(2.4) the pilot process matrix Temp in the i-th iteration operation is calculatedi:
Tempi=D-Ei+μi-1Yi-1;
Wherein, TempiFor the rank matrix of m × 4;
(2.5) to the pilot process matrix TempiCarry out singular value decomposition:
(Ui,∑i,Vi)=svd (Tempi),
Wherein UiFor m × m rank matrix, indicate to the pilot process matrix TempiCarry out the left You Te obtained after singular value decomposition
Levy vector matrix;ViFor 4 × 4 rank matrixes, ViIt indicates to the pilot process matrix TempiIt is obtained after progress singular value decomposition
Right tenth of the twelve Earthly Branches feature matrix;∑iFor the rank matrix of m × 4, ∑iIt indicates to the pilot process matrix TempiCarry out singular value decomposition
The singular value matrix obtained afterwards;Svd indicates to carry out singular value decomposition to matrix;
(2.6) the sparse matrix A in the i-th iteration operation is calculatedi:
Αi=Ui(∑i-μi)Vi T,
Wherein, AiFor the rank matrix of m × 4, the transposition of subscript T representing matrix, μiFor the rank matrix of m × 4, μiEach element be
1/ μ in i-th iterationi-1;
(2.7) the Lagrange multiplier matrix Y in the i-th iteration operation is calculatedi:
Yi=Yi-1+μi-1(D-Ei-Ai),
Wherein, YiFor the rank matrix of m × 4;
The corresponding sound disintegrate-quality relative error ξ of i-th iteration operation is calculatedi=| | D-Ei-Ai||1/||D||1, judgement
ξiWhether ε is greater than: if ξiGreater than ε, then μ is calculatedi=min (μi-1×ρ,μi-1×10-7), it enables i add 1, executes step
(2.3);If ξiLess than or equal to ε, judge whether I is equal to 0, if I is equal to 0, thens follow the steps (2.8);If I is not equal to 0, hold
Row step (2.9);
Wherein, min () expression is minimized;
(2.8) E is enabledS=Ei, calculate the initial value of root-mean-square errorIt enables I add 1, and is calculated
The ε of ε '=10 enables ε take ε ', executes step (2.2);
Wherein, | | | |1The 1- norm of representing matrix, sum () indicate to sum to all elements of column vector in bracket, Ei(:,1)
Expression takes matrix E1First row, Es(:, 1) it indicates to take matrix ESFirst row, length () indicates to seek the member of matrix in bracket
Plain number, ones (length (Ei), 1) indicate element be 1 and order be (length (Ei(:, 1)), 1) × 1 matrix;
(2.9) the sparse matrix E in described i-th this interative computation is calculatediCorresponding root-mean-square error:It calculates
RMSEi-S|;If | RMSEi- S | less than 1, then the ε of ε '=10, enables ε take ε ', and I is enabled to add 1, executes step (2.2);If | RMSEi-
S | it is more than or equal to 1, then the ε of ε '=0.1, obtains the adaptive convergence parameter ε '.
3. the method according to claim 1, wherein the step 3 specifically includes:
(3.1) it initializes: obtaining the corresponding input data matrix D of every continuous infrared image of n frame;K indicates the number of iterations, Ek
Indicate the sparse matrix in kth time interative computation, E0For the matrix and E of m × n rank0In element all 0;AkIndicate kth time
Low-rank matrix in interative computation, A0For the matrix and A of m × n rank0In element all 0;;λ is indicated in intra-frame trunk algorithm
In to the sparse matrix EkThe weight of imparting,YkIndicate the Lagrange multiplier square in the interative computation of kth time
Battle array, Y0=D/max (| | D | |2,λ-1||D||∞);μkIndicate Y in kth time interative computationkPenalty factor, μ0=1.0/ | | D | |2;
ρ is μkRegulatory factor, 0.001≤ρ≤5;
(3.2) k=1 is enabled;
(3.3) the sparse matrix E in the kth time interative computation is calculatedk:
Wherein, EkFor m × n rank matrix, B is m × n rank matrix, each element is λ/μ in matrix Bk-1;
(3.4) to the pilot process matrix TempkCarry out singular value decomposition:
Tempk=D-Ek+μk-1Yk-1;
Wherein, TempkFor m × n rank matrix;
(3.5) to the pilot process matrix TempkCarry out different value decomposition
(Uk,∑k,Vk)=svd (Tempk),
Wherein, wherein UkFor m × m rank matrix, indicate to the pilot process matrix TempkIt is obtained after progress singular value decomposition
Left tenth of the twelve Earthly Branches feature matrix;VkFor n × n rank matrix, VkIt indicates to the pilot process matrix TempkAfter carrying out singular value decomposition
Obtained right tenth of the twelve Earthly Branches feature matrix;∑kFor m × n rank matrix, ∑kIt indicates to the pilot process matrix TempkIt carries out unusual
The singular value matrix that value obtains after decomposing;
(3.6) the sparse matrix A in the kth time interative computation is calculatedk:
Αk=Uk(∑i-μk)Vk T,
Wherein, AkFor m × n rank matrix, μkFor m × n rank matrix, μkEach element be 1/ μ in kth time iterationk-1;
(3.7) the Lagrange multiplier matrix Y in the interative computation of the kth time is calculatedk:
Yk=Yk-1+μk-1(D-Ek-Ak);
Wherein, YkFor m × n rank matrix;
(3.8) the corresponding sound disintegrate-quality relative error ξ of kth this interative computation is calculatedk=| | D-Ek-Ak||1/||D|
|1If ξk> ε ', then be calculated μk=min (μk-1×ρ,μk-1×10-7), it enables k add 1, then executes step (3.2);Otherwise
It executes step (3.9);
(3.9) stop iteration, export the sparse matrix Ek;
Wherein, EkFor m × n rank matrix;
(3.10) sparse matrix is enabledRestore column vectorObtain the corresponding sparse matrix E of each frame infrared image1,E2,…,En;
Wherein,Indicate p-th of column vector, p ∈ [1,2 ..., n],EpOrder
For h × w,
4. according to the method in claim 2 or 3, which is characterized in that the specific generating mode of the input data matrix D
Are as follows:
Infrared image continuous for l frame obtains the corresponding data matrix D of each frame infrared image1,D2,…,Dl;It extracts described every
The even number line of the corresponding data matrix of one frame infrared image and the data of even column obtain adopting under each frame infrared image is corresponding
Sample data matrix D1',D2',…,Dl';The corresponding down-sampled data matrix of each frame infrared image is stretched as according to column
Vector obtains the corresponding column vector of the corresponding down-sampled data matrix of each frame infrared imageAnd then it obtains
Input data matrix
Wherein, l >=4, the element in data matrix are the pixel value of each frame infrared image corresponding position, data matrix D1,
D2,…,DlOrder be H × W;The down-sampled data matrix order isOrder be m × 1,
The order of the input data matrix D is m × l.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811389171.2A CN109584273B (en) | 2018-11-21 | 2018-11-21 | Moving target detection method based on self-adaptive convergence parameters |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811389171.2A CN109584273B (en) | 2018-11-21 | 2018-11-21 | Moving target detection method based on self-adaptive convergence parameters |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109584273A true CN109584273A (en) | 2019-04-05 |
CN109584273B CN109584273B (en) | 2022-11-18 |
Family
ID=65923213
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811389171.2A Active CN109584273B (en) | 2018-11-21 | 2018-11-21 | Moving target detection method based on self-adaptive convergence parameters |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109584273B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113674258A (en) * | 2021-08-26 | 2021-11-19 | 展讯通信(上海)有限公司 | Image processing method and related equipment |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2013143102A (en) * | 2012-01-12 | 2013-07-22 | Nikon Corp | Mobile object detection device, mobile object detection method, and program |
US20140112544A1 (en) * | 2011-06-09 | 2014-04-24 | The HongKongUniversity of Science and Technology. | Image based tracking |
CN104361611A (en) * | 2014-11-18 | 2015-02-18 | 南京信息工程大学 | Group sparsity robust PCA-based moving object detecting method |
CN104867162A (en) * | 2015-05-26 | 2015-08-26 | 南京信息工程大学 | Motion object detection method based on multi-component robustness PCA |
CN105931264A (en) * | 2016-04-14 | 2016-09-07 | 西安电子科技大学 | Sea-surface infrared small object detection method |
-
2018
- 2018-11-21 CN CN201811389171.2A patent/CN109584273B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140112544A1 (en) * | 2011-06-09 | 2014-04-24 | The HongKongUniversity of Science and Technology. | Image based tracking |
JP2013143102A (en) * | 2012-01-12 | 2013-07-22 | Nikon Corp | Mobile object detection device, mobile object detection method, and program |
CN104361611A (en) * | 2014-11-18 | 2015-02-18 | 南京信息工程大学 | Group sparsity robust PCA-based moving object detecting method |
CN104867162A (en) * | 2015-05-26 | 2015-08-26 | 南京信息工程大学 | Motion object detection method based on multi-component robustness PCA |
CN105931264A (en) * | 2016-04-14 | 2016-09-07 | 西安电子科技大学 | Sea-surface infrared small object detection method |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113674258A (en) * | 2021-08-26 | 2021-11-19 | 展讯通信(上海)有限公司 | Image processing method and related equipment |
CN113674258B (en) * | 2021-08-26 | 2022-09-23 | 展讯通信(上海)有限公司 | Image processing method and related equipment |
Also Published As
Publication number | Publication date |
---|---|
CN109584273B (en) | 2022-11-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110991299B (en) | Confrontation sample generation method aiming at face recognition system in physical domain | |
CN109583322A (en) | A kind of recognition of face depth network training method and system | |
CN112285712A (en) | Method for improving detection precision of ship on shore in SAR image | |
CN110135344B (en) | Infrared dim target detection method based on weighted fixed rank representation | |
CN104899866A (en) | Intelligent infrared small target detection method | |
CN111598167B (en) | Small sample image identification method and system based on graph learning | |
CN107274360B (en) | A kind of high spectrum image denoising method based on Fisher dictionary learning, low-rank representation | |
Kyono et al. | Machine learning for quality assessment of ground-based optical images of satellites | |
Miller et al. | A data-constrained algorithm for the emulation of long-range turbulence-degraded video | |
CN106504209A (en) | The iteration of passive millimeter wave radar image weights blind deconvolution method again | |
Liu et al. | Combining deep denoiser and low-rank priors for infrared small target detection | |
Sun et al. | Hyperspectral image denoising via low-rank representation and CNN denoiser | |
Pang et al. | STTM-SFR: Spatial–temporal tensor modeling with saliency filter regularization for infrared small target detection | |
CN109584273A (en) | A kind of moving target detecting method based on adaptive convergence parameter | |
Shen et al. | Comprehensive heading error processing technique using image denoising and tilt-induced error compensation for polarization compass | |
CN108038467B (en) | A kind of sparse face identification method of mirror image in conjunction with thickness level | |
Feng et al. | An adaptive padding correlation filter with group feature fusion for robust visual tracking | |
Bouraffa et al. | Context-aware correlation filter learning toward peak strength for visual tracking | |
Chen et al. | Robust motion blur kernel parameter estimation for star image deblurring | |
CN113591797B (en) | Depth video behavior recognition method | |
Jin et al. | Fusion of remote sensing images based on pyramid decomposition with Baldwinian Clonal Selection Optimization | |
CN114550047A (en) | Behavior rate guided video behavior identification method | |
CN114332623A (en) | Method and system for generating countermeasure sample by utilizing spatial transformation | |
CN110443209A (en) | The infrared small target detection method restored based on matrix | |
CN117238026B (en) | Gesture reconstruction interactive behavior understanding method based on skeleton and image features |
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 |