CN104103042A - Nonconvex compressed sensing image reconstruction method based on local similarity and local selection - Google Patents

Abstract

The invention discloses a nonconvex compressed sensing image reconstruction method based on local similarity and local selection. The method comprises the following steps: 1) carrying out observation and reception after an image is partitioned; 2)utilizing a local growth method to carry out clustering on observation vectors of all the image blocks; 3) carrying out population initialization on the image block corresponding to each kind of observation vector according to the scheme that the polyatom direction and monatom direction coexist; 4) utilizing an improved genetic algorithm to carry out crossing, variation and selection operation based on a local selection mechanism on the populations obtained in the step 3), reconstituting corresponding image blocks and obtaining optimal atom combinations; 5) utilizing a clone selection optimization algorithm to study the optimal atom combinations on the aspects of dimension and displacement; and 6) piecing the image blocks obtained in the step 5) together in sequence to obtain a complete reconstructed image, and outputting the complete reconstructed image. The reconstructed image is good in visual effect and high in peak signal to noise ratio, and can be used for nonconvex compressed sensing reconstruction of image signals under the condition of low sampling rate.

Description

A kind of based on local similarity and the local non-protruding compressed sensing image reconstructing method of selecting

Technical field

The invention belongs to technical field of image processing, further relate to Image Reconstruction, specifically a kind of based on local similarity and the local non-protruding compressed sensing image reconstructing method of selecting.

Background technology

In image reconstruction technique field, a kind of new data acquisition theory---compressive sensing theory is a great change of field of information processing in recent years.This theory points out, signal can carry out low speed sampling and a small amount of sampling, and can Accurate Reconstruction, the complexity that greatly reduces like this device storage restriction and calculate.Compressed sensing has become the focus of academia's research at present, and is constantly used in compression imaging system and bio-sensing field.Compressed sensing technology relates generally to the content of following three aspects:: the rarefaction representation of signal, the design of observing matrix and signal reconstruction.Wherein, signal reconstruction is key and the core of compressed sensing technology.

In compressed sensing technology, the restructuring procedure of picture signal be unable to do without solving of underdetermined system of equations problem.E。The people such as Candes prove, if signal is sparse or compressible, the problem that solves the underdetermined system of equations can be converted into solve and minimize l ₀norm problem, thereby reconstruction signal.This source problem of compressed sensing reconstruct is l ₀non-protruding optimization problem under norm.Direct solution l at present ₀the method of norm problem is to shrink threshold algorithm two classes of IHT as representative taking orthogonal matching pursuit OMP algorithm as the greedy algorithm of representative with taking iteration threshold.

OMP algorithm is in the process of each iteration, and the thought based on greedy the means by local optimum are selected a atom that can matched signal structure, and approach through a series of methods structure the sparse of signal that progressively increase progressively.But OMP algorithm can not all be realized Accurate Reconstruction to all picture signals, reconstruction result is not very accurate, and algorithm does not have robustness yet.

IHT algorithm is also based on l ₀the reconstructing method of norm, the method is carried out the random observation of low sampling and uses orthogonal basis sparse signal, and then by selecting important rarefaction representation coefficient, gives up unessential rarefaction representation coefficient and carry out reconstructed image.The shortcoming of IHT algorithm is to measuring the undue dependence of matrix, and computation complexity is high, and operation time is long, and the size of threshold value is larger on the reconstruction result impact of picture signal.

Patented claim " based on redundant dictionary and the sparse non-protruding compressed sensing image reconstructing method of the structure " (publication number: CN103295198A of Xian Electronics Science and Technology University, application number: CN 201310192104, the applying date: on May 13rd, 2013) in a kind of image compression reconstruction method based on non-convex model is disclosed, call patented method in the following text, the method is used mutual neighbour's technology to carry out cluster to observation vector, according to the difference between observation vector (being Euclidean distance), adjacent or non-conterminous observation vector can be gathered is a class; According to single direction initialization population, use traditional genetic algorithm to find out in dictionary direction preferably former sub-portfolio for each class observation vector, preserve population; Be that it is in the former sub-portfolio of determining to find out in direction optimum in yardstick and displacement to using clonal selection algorithm after each image block execution population extended operation; Image block is reconstructed by the former sub-portfolio of optimum.This inventive method reconstruct effect compared with OMP, IHT method increases, and experiment shows, in the time that sampling rate 30% is above, can realize the Accurate Reconstruction to image.But along with sampling rate reduces, observation vector carry information reduces, and directly carries out cluster according to the Euclidean distance between observation vector, easily occurs more cluster mistake.In addition in this invention population according to single direction initialization, the method has good reconstruct effect for the single texture image piece of direction, but smooth image piece direction is various, therefore, in the situation that sampling rate reduces, the reconstruct effect of image smoothing part is undesirable.Meanwhile, this invention utilizes traditional genetic algorithm to learn, and its individual choice method easily causes evolution result precocity, is absorbed in local optimum, thereby affects results of learning.Therefore the Image Reconstruction effect of this invention under low sampling rate is not ideal enough.

In sum, existing OMP, IHT, based on the methods such as redundant dictionary and the sparse non-protruding compressed sensing image reconstructing method of structure in sampling rate lower than 30% in the situation that, existing based on minimizing l ₀the compressed sensing restructing algorithm of norm is all not accurate enough to the reconstruct of image, therefore also needs explore further and study.

Summary of the invention

The present invention is directed to above-mentioned prior art under low sampling rate to not accurate enough this problem of Image Reconstruction, propose a kind ofly based on local similarity and the local non-protruding compressed sensing image reconstructing method of selecting, can not obtain good visual effect, reconstructed image that Y-PSNR PSNR is higher for solving prior art.

Technical scheme of the present invention is: a kind of based on local similarity and the local non-protruding compressed sensing image reconstructing method of selecting, comprise the steps:

(1) take original image, and it is carried out observing and receiving after piecemeal;

(2) calculate the standard deviation of each observation vector and utilize the local similarity of observation vector, adopting the method for local growth to carry out cluster to the observation vector of all image blocks;

(3) to image block corresponding to each class observation vector by polyatom direction and monatomic direction the scheme initialization population of depositing;

(4) utilize improved evolutionary programming algorithm that the population in step (3) is intersected, made a variation and select machine-processed selection operation based on part, image block corresponding to the every class observation vector of reconstruct, obtains former sub-portfolio optimum in dictionary direction;

(5) recycling Immune Clone Selection optimized algorithm is learnt out former sub-portfolio optimum in yardstick and displacement;

(6) reconstructed image the piece corresponding all observation vectors that obtain in step 5 is stitched together according to the order of sequence and obtains view picture reconstructed image, output view picture reconstructed image.

Wherein the concrete steps of step (2) are as follows:

(2.1) calculate the standard deviation of each observation vector;

(2.2) all image blocks are all arranged to a cluster mark, be initially all labeled as 0, wherein, mark 0 represents not by cluster, and mark 1 represents to be included in a certain classification;

(2.3), from first image block, successively each image block is done to following operation: if image block cluster is labeled as 1, do not operate, if be labeled as 0, carry out M _ithe cluster of class, clustering method is as follows:

(2.3.1) taking current image block i as seed image block, observation vector corresponding seed image block is added to class M _iand as Seeding vector;

(2.3.2) by the standard deviation of Seeding vector respectively with seed image block eight adjacent image piece A around ₁, A ₂a ₈the standard deviation of corresponding observation vector is subtracted each other and is obtained C ₁, C ₂c ₈; If image block A _i(i=1,2 ... 8) cluster be labeled as 0, and | C _i|≤τ (i=1,2 ... 8), wherein τ is threshold value, by image block A _icluster mark is set to 1, and by its corresponding observation vector y _iadd class M _iin; The observation vector that these are added is according to its respective standard difference | C _i| the order increasing progressively is at class M _imiddle arrangement;

(2.3.3) at class M _iin, if Seeding vector is last element, M _iclass cluster completes; Otherwise, make M _ifirst element after middle Seeding vector is new Seeding vector, making image block corresponding to new seed vector is new seed image block, upgrade threshold tau=max (0.1, τ-0.1), then repeating step (2.3.2) and (2.3.3).

Wherein the related initialization of population scheme of step (3) is as follows:

If the population of the image block that such observation vector is corresponding is A, in population, individual amount is (H+P); Again by following Optimization Steps:

(3.1) individual for front H, utilizing sliding window method is that each individuality is chosen 10 directions, and sliding window scope is 1 to P, and length of window is 10, and sliding window lap is 8; From corresponding each the sub-dictionary of these 10 directions, k atom of random generation forms this individuality;

(3.2) for below P individual, make each individuality comprise successively a direction in 1 to P, produce k atom and form this individuality at random from sub-dictionary corresponding to this direction.

Wherein the operation steps of above-mentioned steps (4) is as follows:

(4.1) calculate the fitness of all individualities in the population of the image block that each class observation vector is corresponding according to following fitness function:

$f\left(X_m\right)=\frac{1}{\underset{i=1}{\overset{j}{\mathrm{\Σ}}}{\left|\right|y_i-\mathrm{\Φdec}\left(X_m\right){\mathrm{\α}}_i\left|\right|}_2^2}$

Wherein, f (X _m) be m individual fitness value in the population A of the image block that such observation vector is corresponding, the label that i is observation vector, j is the sum of observation vector in each class after cluster, y _ifor i observation vector in class, Φ is Gauss's observing matrix, X _mm individual all gene position in population A, dec (X _m) representing these gene position corresponding former sub-portfolio in dictionary, this former sub-portfolio is exactly a sub-dictionary, α _ifor the rarefaction representation coefficient vector of image block corresponding to i observation vector in class, it is by the generalized inverse matrix of sensing matrix and this observation vector y _imultiply each other and obtain, sensing matrix is by Gauss's observing matrix Φ and sub-dictionary dec (X _m) multiplying each other obtains, be vectorial two norms square.

(4.2) population A of image block corresponding to such observation vector is carried out to the intersection of improved genetic algorithms method, variation and locally select operation, upgrades and preserves the population A after operating ₁;

(4.2.1) interlace operation: current population A is divided into two sub-populations, and first sub-population Q1 is made up of front H multi-direction individuality, second sub-population Q2 is made up of P one direction individuality below.For each individual i in sub-population Q1 prepares a set U _i, this set is initialized as that { i} is random one (0,1) the interval numerical value v that generates of individual i _iif, numerical value v _ibe less than crossover probability p _c, random another individuality of selection from sub-population Q1, and select at random a crossover location, these two individualities are carried out to single-point intersection and obtain two new individualities, add these two new individualities to set U _iin; In like manner, for each individual i in sub-population Q2 prepares a set V _i, this set is initialized as that { i} is random one (0,1) the interval numerical value v that generates of individual i _iif, numerical value v _ibe less than crossover probability p _c, random another individuality of selection from sub-population Q2, and select at random a crossover location, these two individualities are carried out to single-point intersection and obtain two new individualities, add these two new individualities to set V _iin;

(4.2.2) mutation operation: generate one (0,1) interval numerical value u for each individual i in sub-population Q1 is random _iif this numerical value is less than variation Probability p _m, to variation position of the random selection of this individuality, and select at random a numerical value that is not more than dictionary scale to replace the numerical value of variation position, obtain a new individuality, this individuality is added to the set U of its correspondence _iin; In like manner, generate one (0,1) interval numerical value u for each individual i in sub-population Q2 is random _iif this numerical value is less than variation Probability p _m, to variation position of the random selection of this individuality, and select at random a numerical value that is not more than dictionary scale to replace the numerical value of variation position, obtain a new individuality, this individuality is added to the set V of its correspondence _iin;

(4.2.3) local selection operates: be set U corresponding to the each individual i of sub-population Q1 _iin select the individuality of fitness minimum, with it replace the individual i in former population A; In like manner, be set V corresponding to the each individual i of sub-population Q2 _iin select the individuality of fitness minimum, replace the individual i in former population A with it, obtain new population A ₁.

(4.3) judge population A ₁whether meet the iteration stopping condition of genetic algorithm, if meet, go to step (4.4), if do not meet, by population A ₁as the population A of image block corresponding to such observation vector, go to step (4.1).

(4.4) individuality that selection fitness is the highest is as optimum individual, one group of corresponding with optimum individual the rarefaction representation coefficient vector of image block corresponding such each observation vector Ridgelet base atom is multiplied each other, obtain the reconstructed image piece that such observation vector is corresponding, preserve the population A after optimization renewal in step (4.2) ₂.

(4.5) population of image block corresponding to all class observation vectors is performed step to (4.1) successively to step (4.4), obtain the reconstructed image piece that all class observation vectors are corresponding and optimize the population after upgrading;

Described step (5) operation steps is as follows:

(5.1) for the each reconstructed image piece B obtaining in step 4, by its corresponding population A of preserving in step (4.4) ₂as its initialization population;

(5.2) add two optimum antibody corresponding 8 adjacent image pieces of image block B to initialization population A ₂, and remove and repeat to obtain population A after antibody ₃, calculate population A ₃scale, be designated as l;

(5.3) first, calculate population A corresponding to this image block B according to following affinity function ₃in the affinity of all antibody:

$g\left(X_m\right)=\frac{1}{{\left|\right|y_0-\mathrm{\Φdec}\left(X_m\right){\mathrm{\α}}_m\left|\right|}_2^2}$

Wherein, g (X _m) be population A corresponding to this image block B ₃in the affinity value of m antibody, y ₀be the observation vector of image block B, Φ is Gauss's observing matrix, X _mall gene position of m antibody in population A, dec (X _m) representing these gene position corresponding former sub-portfolio in dictionary, this former sub-portfolio is exactly a sub-dictionary, α _mfor image block B is at sub-dictionary dec (X _m) under rarefaction representation coefficient vector, it is by the generalized inverse matrix of sensing matrix and observation vector y ₀multiply each other and obtain, sensing matrix is by Gauss's observing matrix Φ and sub-dictionary dec (X _m) multiplying each other obtains, be vectorial two norms square;

(5.4) population A corresponding to image block B again ₃carry out the clone of clonal selection algorithm, variation, selects operation, upgrades and preserves the population A after operating ₄;

(5.5) judge whether to meet iteration stopping condition, if meet, go to step (5.6), if do not meet, by the population A of upgrading after operation ₄as population A corresponding to image block B ₃, go to step (5.3).

(5.6) antibody that selection affinity is the highest, as optimum antibody, multiplies each other one group of corresponding with optimum antibody the rarefaction representation coefficient vector of image block B Ridgelet base atom, obtains the reconstructed image piece of image block B, preserves the population A of optimizing after upgrading ₅.

(5.7) population of all observation vector correspondence image pieces is performed step to (5.1) successively to step (5.6), obtain the reconstructed image piece that all observation vectors are corresponding and optimize the population after upgrading.

Step 6, is stitched together reconstructed image the piece corresponding all observation vectors that obtain in step (5.7) according to the order of sequence and obtains view picture reconstructed image, output view picture reconstructed image.

Such scheme is main thought of the present invention, comprises image is carried out to piecemeal observation and reception, uses the clustering algorithm based on observation vector local similarity and local growth to carry out cluster to the observation vector of all image blocks.According to polyatom direction and monatomic direction the scheme of depositing, each class observation vector is carried out to initialization of population, use and select machine-processed improved genetic algorithms method to solve the more excellent former sub-portfolio of such image block in dictionary direction based on part; To each image block, make the population obtaining in last step as initialization population, after this population is carried out to population extended operation removal repetition antibody, determine population scale, use clonal selection algorithm further to search for more excellent former sub-portfolio with regard to yardstick and displacement parameter on this population scale.By above step, each image block will obtain one group of preferably base atom, and then can obtain the reconstruction result of each image block.All reconstructed image pieces are spelled to the reconstruction result that can obtain entire image.

The present invention compared with prior art has the following advantages:

1. the present invention takes full advantage of the direction character of image and the direction character of Ridgelet dictionary atom.In Ridgelet redundant dictionary, atom is by direction, yardstick and displacement parameter decision, the direction of its Atom is even more important to the adaptive rarefaction representation of image block.Experiment shows, the atom that texture image piece is carried out to rarefaction representation has the concentrated feature of directivity, and the atom that smooth image piece is carried out to rarefaction representation has the various and homodisperse feature of direction.Under low sampling rate, because the information that observation vector carries is less, cannot judge and treat that reconstructed image piece is smooth or texture block according to prior imformation, therefore, the present invention carries out initialization of population to the similar image block of each class according to the scheme of the atomic orientation feature of taking into account smooth and texture block, makes to exist in population the individuality that has multiple directions and single direction simultaneously.And the patented method of mentioning is above according to single direction initialization population, poor for the smooth image piece reconstruct effect that atomic orientation is various, therefore, the present invention has solved effectively under low sampling rate, the problem that reconstructed image visual effect is poor, Y-PSNR is lower that existing compressed sensing reconfiguration technique obtains, make the present invention compared with the conventional method, improved visual effect and the peak value to-noise ratio of reconstructed image.

2. the present invention takes full advantage of the local similar feature of natural image, uses the method based on observation vector local similarity and local growth to carry out cluster to observation vector.Observe and experience show, adjacent image piece has higher similarity, further found through experiments, the observation vector that similar image piece is corresponding is also very approaching, therefore can carry out cluster by observation vector standard deviation.In order to improve efficiency and the piece constraining force of evolutionary learning, adopt the mode of local growth to carry out cluster.And mutual neighbour's clustering method of patented method carries out cluster based on whole observation vectors, according to the Euclidean distance between observation vector, adjacent and non-conterminous observation vector being gathered is a class, in sampling rate reduction situation, observation vector carry information reduces, directly certainly will bring mistake according to the Euclidean distance cluster between observation vector, and cluster based on whole observation vectors can increase the number of this mistake.The present invention is according to observation vector standard deviation cluster, and according to the local similarity of the image cluster of only growing in subrange, can greatly reduce cluster mistake, therefore the invention solves the problem at low sampling rate hypograph reconstruct weak effect, make the present invention improve the reconstruction quality of low sampling rate image, obtain the higher reconstructed image of Y-PSNR.

3. utilization of the present invention is selected machine-processed improved genetic algorithms method based on part, has ensured the diversity of population direction.The selection algorithm using in patented method, is to select in all individualities of population, the individuality with some direction may be got rid of too early, thereby be caused the precocity of evolving, and is absorbed in local optimum.The present invention adopts the local mechanism of selecting, in the individuality individual from population, this individuality interlace operation produces, the individuality that this individual variation operation produces, select more excellent individuality to add new population, can avoid like this population diversity to lose, thereby improve the accuracy of Image Reconstruction, therefore the invention solves under low sampling rate, the problem that reconstructed image visual effect is poor, Y-PSNR is lower that existing compressed sensing reconfiguration technique obtains, make the present invention compared with the conventional method, improved visual effect and the peak value to-noise ratio of reconstructed image.

4. in the present invention, carry out population extended operation after genetic evolution study finishes time, that two optimum antibody of 8 adjacent pieces around image block are all added to population, an instead of antibody in patent, in addition, algorithm is not fixed unified population scale in the time using Immune Clone Selection study herein, but expanding population and removing and repeat after antibody, determine the population scale of current image block according to antibody number in current population, therefore the population scale that each image block may be corresponding different, and this scale is more than or equal to the unified scale of formulating in patented method, thereby be more conducive to population diversity, for further searching for more selection be provided, make the present invention compared with the conventional method, visual effect and the peak value to-noise ratio of reconstructed image are improved.

In brief, the present invention is using genetic algorithm and clonal selection algorithm as nonlinear optimization reconstructing method, mechanism is selected in the direction character, the direction character of natural image and the part of local similarity and genetic algorithm that take full advantage of Ridgelet dictionary atom, and designed corresponding operator, finally improve the reconstruction quality of low sampling rate image.

Brief description of the drawings

Below in conjunction with embodiment accompanying drawing, the invention will be further described

Fig. 1 is Image Reconstruction FB(flow block) of the present invention;

Fig. 2 is standard testing image Barbara figure, Lena figure and their partial enlarged drawing;

Fig. 3 is OMP method and IHT method and the patented method reconstruct effect contrast figure to standard testing image Barbara figure under 25% sampling rate by the present invention and prior art;

Fig. 4 is OMP method and IHT method and the patented method reconstruct effect contrast figure to standard testing image Lena figure under 25% sampling rate by the present invention and prior art;

Fig. 5 is OMP method and IHT method and the patented method trend comparison diagram that the Y-PSNR PSNR of the reconstructed image to Barbara figure, Lena figure, Peppers figure changes with sampling rate respectively of the present invention and prior art.

Fig. 6 be respectively under 20% and 30% sampling rate to observation vector standard deviation statistics between adjacent image piece observation vector standard deviation statistics, similar adjacent smooth interblock observation vector standard deviation statistics and similar texture block in standard testing image Barbara figure.

Embodiment

Below in conjunction with 1-6 accompanying drawing, the present invention is described further.

Embodiment 1, describes in detail with reference to Fig. 1.

The present invention is a kind of based on local similarity and the local non-protruding compressed sensing image reconstructing method of selecting, the method can be carried out low speed sampling and a small amount of sampling to picture signal, and then Accurate Reconstruction image, the complexity that greatly reduces like this device storage restriction and calculate, concrete implementation step is as follows:

Step (1), carries out observing and receiving after piecemeal to original image.

Input original image is also divided into the not overlapping block of 16*16, utilizes random Gaussian observing matrix Φ respectively each piece observe and obtained measuring vectorial y, and transmitting terminal sends the vectorial y of measurement of observing matrix Φ and each piece, and receiving end receives;

In the present embodiment, 512 × 512 image is divided into 16 × 16 image block, obtains 1024 image blocks; In computing machine, with matlab software, all image blocks are preserved into column vector, column vector corresponding all image blocks is multiplied each other with identical Gauss's observing matrix, obtain 1024 observation vectors.

Step (2) is calculated the standard deviation of each observation vector and is utilized the local similarity of observation vector, adopts the method for local growth to carry out cluster to the observation vector of all image blocks;

Step (3) to image block corresponding to each class observation vector by polyatom direction and monatomic direction the scheme initialization population of depositing;

Step (4) is utilized improved evolutionary programming algorithm that the population in step (3) is intersected, made a variation and is selected machine-processed selection operation based on part, and image block corresponding to the every class observation vector of reconstruct, obtains former sub-portfolio optimum in dictionary direction;

Step (5) recycling Immune Clone Selection optimized algorithm is learnt out former sub-portfolio optimum in yardstick and displacement;

Reconstructed image the piece corresponding all observation vectors that obtain in step 5 is stitched together according to the order of sequence and obtains view picture reconstructed image, output view picture reconstructed image.

This scheme has solved effectively under low sampling rate, the problem that reconstructed image visual effect is poor, Y-PSNR is lower that existing compressed sensing reconfiguration technique obtains, make the present invention compared with the conventional method, improved visual effect and the peak value to-noise ratio of reconstructed image.

Embodiment 2,1-6 describes by reference to the accompanying drawings.

On the basis of embodiment 1, described step (2), utilizes the local similarity of observation vector standard deviation, uses the method for local growth to carry out cluster to the observation vector of all image blocks, specifically comprises step:

2.1) calculate the standard deviation of each observation vector.

2.2) all image blocks are all arranged to a cluster mark, be initially all labeled as 0, wherein, mark 0 represents not by cluster, and mark 1 represents to be included in a certain classification.

2.3), from first image block, successively each image block is done to following operation: if image block cluster is labeled as 1, do not operate, if be labeled as 0, carry out M _ithe cluster of class.

Carry out M _ithe concrete steps of class cluster are as follows:

The first step, taking current image block i as seed image block, adds class M by observation vector corresponding seed image block _iand as Seeding vector;

Second step, by the standard deviation of Seeding vector respectively with seed image block eight adjacent image piece A around ₁, A ₂a ₈the standard deviation of corresponding observation vector is subtracted each other and is obtained C ₁, C ₂c ₈; If image block A _i(i=1,2 ... 8) cluster be labeled as 0, and | C _i|≤τ (i=1,2 ... 8), wherein τ is threshold value, by image block A _icluster mark is set to 1, and by its corresponding observation vector y _iadd class M _iin; The observation vector that these are added is according to its respective standard difference | C _i| the order increasing progressively is at class M _imiddle arrangement;

The 3rd step, at class M _iin, if Seeding vector is last element, M _iclass cluster completes; Otherwise, make M _ifirst element after middle Seeding vector is new Seeding vector, and making image block corresponding to new seed vector is new seed image block, upgrades threshold tau=max (0.1, τ-0.1), then repeats second step and the 3rd step.

According to experimental result, be below 20% time in sampling rate, τ initial value is made as 0.3; Be 20% when above in sampling rate, τ initial value is made as 0.4.

Described step (3), to image block corresponding to each class observation vector by polyatom direction and monatomic direction the scheme initialization population of depositing.Concrete steps are as follows:

3.1) generate Ridgelet redundant dictionary,, successively to the atom numbering in dictionary atom identical direction is organized together and obtains a sub-dictionary since 1, the direction number of dictionary Atom is P, obtains altogether the sub-dictionary of P different directions; Degree of rarefication value by image block under Ridgelet redundant dictionary is set as k, k be one with image-related positive integer.

In the present embodiment, in Ridgelet redundant dictionary, there are 4201 base atoms, 4201 base atoms are numbered respectively to 1,2,3 ..., 4201, one have the sub-dictionary of 36 different directions, and the degree of rarefication of all image blocks is set as 32.

3.2), according to the population A of image block corresponding to such observation vector of atomic orientation initialization, in population, individual amount is (H+P).

In the present embodiment, H is 14.The initialization step of population A is as follows:

The first step, for front 14 individualities, utilizing sliding window method is that each individuality is chosen 10 directions, and sliding window scope is 1 to 36, and length of window is 10, and sliding window lap is 8; From corresponding each the sub-dictionary of these 10 directions, 32 atoms of random generation form this individuality; .

Second step, for 36 individualities below, makes each individuality comprise successively a direction in 1 to 36, produces 32 atoms and forms this individuality at random from sub-dictionary corresponding to this direction;

Described step (4), utilizes improved evolutionary programming algorithm that the population in step (3) is intersected, made a variation and selects machine-processed selection operation based on part, and image block corresponding to the every class observation vector of reconstruct, obtains former sub-portfolio optimum in dictionary direction; Concrete steps are as follows:

4.1) calculate the fitness of all individualities in the population of the image block that each class observation vector is corresponding according to following fitness function:

$f\left(X_m\right)=\frac{1}{\underset{i=1}{\overset{j}{\mathrm{\Σ}}}{\left|\right|y_i-\mathrm{\Φdec}\left(X_m\right){\mathrm{\α}}_i\left|\right|}_2^2}$

Wherein, f (X _m) be m individual fitness value in the population A of the image block that such observation vector is corresponding, the label that i is observation vector, j is the sum of observation vector in each class after cluster, y _ifor i observation vector in class, Φ is Gauss's observing matrix, X _mm individual all gene position in population A, dec (X _m) representing these gene position corresponding former sub-portfolio in dictionary, this former sub-portfolio is exactly a sub-dictionary, α _ifor the rarefaction representation coefficient vector of image block corresponding to i observation vector in class, it is by the generalized inverse matrix of sensing matrix and this observation vector y _imultiply each other and obtain, sensing matrix is by Gauss's observing matrix Φ and sub-dictionary dec (X _m) multiplying each other obtains, be vectorial two norms square.

4.2) population A of image block corresponding to such observation vector is carried out to the intersection of genetic algorithm, variation, selects operation, upgrades and preserves the population A after operating ₁;

Intersect, variation, select the concrete steps of operation as follows:

The first step, interlace operation: current population A is divided into two sub-populations, and first sub-population Q1 is made up of front 14 multi-direction individualities, second sub-population Q2 is made up of 36 one direction individualities below.For each individual i in sub-population Q1 prepares a set U _i, this set is initialized as that { i} is random one (0,1) the interval numerical value v that generates of individual i _iif, numerical value v _ibe less than crossover probability p _c, random another individuality of selection from sub-population Q1, and select at random a crossover location, these two individualities are carried out to single-point intersection and obtain two new individualities, add these two new individualities to set U _iin; In like manner, for each individual i in sub-population Q2 prepares a set V _i, this set is initialized as that { i} is random one (0,1) the interval numerical value v that generates of individual i _iif, numerical value v _ibe less than crossover probability p _c, random another individuality of selection from sub-population Q2, and select at random a crossover location, these two individualities are carried out to single-point intersection and obtain two new individualities, add these two new individualities to set V _iin.

Second step, mutation operation: generate one (0,1) interval numerical value u for each individual i in sub-population Q1 is random _iif this numerical value is less than variation Probability p _m, to variation position of the random selection of this individuality, and select at random a numerical value that is not more than dictionary scale to replace the numerical value of variation position, obtain a new individuality, this individuality is added to the set U of its correspondence _iin; In like manner, generate one (0,1) interval numerical value u for each individual i in sub-population Q2 is random _iif this numerical value is less than variation Probability p _m, to variation position of the random selection of this individuality, and select at random a numerical value that is not more than dictionary scale to replace the numerical value of variation position, obtain a new individuality, this individuality is added to the set V of its correspondence _iin;

The 3rd step, local selection operates: be set U corresponding to the each individual i of sub-population Q1 _iin select the individuality of fitness minimum, with it replace the individual i in former population A; In like manner, be set V corresponding to the each individual i of sub-population Q2 _iin select the individuality of fitness minimum, replace the individual i in former population A with it, obtain new population A ₁.

4.3) judge population A ₁whether meet the iteration stopping condition of genetic algorithm, if meet, go to step 4.4), if do not meet, by population A ₁as the population A of image block corresponding to such observation vector, go to step 4.1).

4.4) individuality that selection fitness is the highest is as optimum individual, one group of corresponding with optimum individual the rarefaction representation coefficient vector of image block corresponding such each observation vector Ridgelet base atom is multiplied each other, obtain the reconstructed image piece that such observation vector is corresponding, preserve step 4.2) the middle population A of optimizing after upgrading ₂.

4.5) population of image block corresponding to all class observation vectors is performed step to 4.1 successively) to step 4.4), obtain the reconstructed image piece that all class observation vectors are corresponding and optimize the population after upgrading;

Described step (5), utilizes Immune Clone Selection optimized algorithm to learn out former sub-portfolio optimum in yardstick and displacement.Concrete steps are as follows:

5.1) for the each reconstructed image piece B obtaining in step 4, it is corresponding at step 4d) population A of preserving ₂as its initialization population.

5.2) add two optimum antibody corresponding 8 adjacent image pieces of image block B to initialization population A ₂, and remove and repeat to obtain population A after antibody ₃, calculate population A ₃scale, be designated as l;

5.3) calculate population A corresponding to this image block B according to following affinity function ₃in the affinity of all antibody:

$g\left(X_m\right)=\frac{1}{{\left|\right|y_0-\mathrm{\Φdec}\left(X_m\right){\mathrm{\α}}_m\left|\right|}_2^2}$

Wherein, g (X _m) be population A corresponding to this image block B ₃in the affinity value of m antibody, y ₀be the observation vector of image block B, Φ is Gauss's observing matrix, X _mall gene position of m antibody in population A, dec (X _m) representing these gene position corresponding former sub-portfolio in dictionary, this former sub-portfolio is exactly a sub-dictionary, α _mfor image block B is at sub-dictionary dec (X _m) under rarefaction representation coefficient vector, it is by the generalized inverse matrix of sensing matrix and observation vector y ₀multiply each other and obtain, sensing matrix is by Gauss's observing matrix Φ and sub-dictionary dec (X _m) multiplying each other obtains, be vectorial two norms square.

5.4) population A corresponding to image block B ₃carry out the clone of clonal selection algorithm, variation, selects operation, upgrades and preserves the population A after operating ₄.

The clone of clonal selection algorithm, variation, select the concrete steps of operation as follows:

The first step, will population A be cloned ₃in each antibody copy 10, add population A to by copying the antibody obtaining ₃in, obtain the population after clone operations;

Second step, to random one (0,1) the interval numerical value that generates of the each antibody in the population after clone operations, if this numerical value is less than variation Probability p _m', on this antibody, random selection, selects at random one to replace into from the atom of the different displacements of the variation equidirectional same yardstick of position atom at a variation position from dictionary atom set, obtains a new antibodies, this antibody is added in population, obtain the population after mutation operation;

The 3rd step, the antibody that in the population from mutation operation, front l the affinity value of selection is higher is as population A ₄, wherein l is by initial population A ₃make out the scale.

5.5) judge whether to meet iteration stopping condition, if meet, go to step 5.6), if do not meet, by the population A of upgrading after operation ₄as population A corresponding to image block B ₃, go to step 5.3).

5.6) antibody that selection affinity is the highest, as optimum antibody, multiplies each other one group of corresponding with optimum antibody the rarefaction representation coefficient vector of image block B Ridgelet base atom, obtains the reconstructed image piece of image block B, preserves the population A of optimizing after upgrading ₅.

5.7) population of all observation vector correspondence image pieces is performed step to 5.1 successively) to step 5.6), obtain the reconstructed image piece that all observation vectors are corresponding and optimize the population after upgrading.

Described step (6), is stitched together all reconstructed image pieces according to the order of sequence and obtains view picture reconstructed image, because the reconstruction quality of each image block has obtained optimization, so the reconstruction result of entire image also will be more accurate, and output view picture reconstructed image.

Effect of the present invention can further illustrate by following emulation experiment.

1. simulated conditions

(1) that this experiment is used is Barbara figure, the Lena figure in 512 × 512 standard testing image storehouse, and the size of image block is decided to be 16 × 16;

(2) this experimental observation matrix is random Gaussian observing matrix, and sampling rate is 20%, 25%, 30%;

(3) the Ridgelet redundant dictionary scale that this experiment adopts is 4201, has 36 directions;

(4) degree of rarefication of this experimental image piece is set as fixed value 32;

(5) Population Size of this experiment is 50;

(6) crossover probability of this experiment genetic algorithm is 0.6, and variation probability is 0.01; The variation probability of clonal selection algorithm is 0.3;

(7) this experiment genetic algorithm iteration 100 times, clonal selection algorithm iteration 20 times.

2. emulation content and result

(1) emulation 1

In this emulation, checking the present invention has better reconstruct effect than the OMP method of prior art, IHT method and the non-protruding compressed sensing image reconstructing method based on redundant dictionary in visual effect.Experimental result is as Fig. 2, shown in Fig. 3 and Fig. 4, wherein Fig. 2 (a1) is Barbara test primitive figure, Fig. 2 (a2) is the partial enlarged drawing of Fig. 2 (a1), Fig. 3 (a1) is the restructuring graph of the present invention in the time that sampling rate is 25%, Fig. 3 (a2) is the partial enlarged drawing of Fig. 3 (a1), Fig. 3 (b1) is the restructuring graph of OMP method in the time that sampling rate is 25%, Fig. 3 (b2) is the partial enlarged drawing of Fig. 3 (b1), Fig. 3 (c1) is the restructuring graph of IHT method in the time that sampling rate is 25%, Fig. 3 (c2) is the partial enlarged drawing of Fig. 3 (c1), Fig. 3 (d1) is the restructuring graph of patented method in the time that sampling rate is 25%, Fig. 3 (d2) is the partial enlarged drawing of Fig. 3 (d1), Fig. 2 (b1) is Lena test primitive figure, Fig. 2 (b2) is the partial enlarged drawing of Fig. 2 (b1), Fig. 4 (a1) is the restructuring graph of the present invention in the time that sampling rate is 25%, Fig. 4 (a2) is the partial enlarged drawing of Fig. 4 (a1), Fig. 4 (b1) is the restructuring graph of OMP method in the time that sampling rate is 25%, Fig. 4 (b2) is the partial enlarged drawing of Fig. 4 (b1), Fig. 4 (c1) is the restructuring graph of IHT method in the time that sampling rate is 25%, Fig. 4 (c2) is the partial enlarged drawing of Fig. 4 (c1), Fig. 4 (d1) is the restructuring graph of patented method in the time that sampling rate is 25%, Fig. 4 (d2) is the partial enlarged drawing of Fig. 4 (d1).

From Fig. 3 (b1), in Fig. 3 (c1) and Fig. 3 (d1), can find out, OMP method, image ratio after IHT method and patented method reconstruct is fuzzyyer, further from partial enlarged drawing 3(b2) and Fig. 3 (c2) can find out, there is smooth block in scarf texture part, there is no better to react the even grain feature of the specific direction that original image has, face is also fuzzyyer, from Fig. 3 (d2), can find out that the texture of scarf and background is mixed and disorderly, from Fig. 3 (a1), can find out, reconstructed image of the present invention is more clear, from Fig. 3 (a2), can find out, compared with above-mentioned three kinds of methods, the mixed and disorderly texture of scarf and background is fewer, it is comparatively accurate that grain direction is recovered, face is also more clear.From Fig. 4 (b1), in Fig. 4 (c1) and Fig. 4 (d1), can find out, OMP method, image ratio after IHT method and patented method reconstruct is fuzzyyer, further from partial enlarged drawing 4(b2) and Fig. 4 (c2) can find out, face, nose and shoulder part are clear not level and smooth, eyes and the brim of a hat feather are fuzzyyer, find out from Fig. 4 (d2), nose, shoulder is level and smooth not, eyes recover not accurate enoughly, from Fig. 4 (a1), can find out, reconstructed image of the present invention is more clear, from Fig. 4 (a2), can find out, compared with above-mentioned three kinds of methods, hair, the recovery effects of shoulder and face is more clear accurately.Therefore,, under identical sampling rate, the visual effect of the image of reconstruct of the present invention is better than OMP method, IHT method and patented method.

(2) emulation 2

In this emulation, checking the present invention has better reconstruct effect than the OMP method of prior art and IHT method on Y-PSNR PSNR.Experimental result as shown in Figure 5, the trend comparison diagram that wherein Fig. 5 (a) changes with sampling rate with the Y-PSNR of the present invention, OMP method, IHT method and patented method reconstructed image respectively for Barbara figure, the trend comparison diagram that wherein Fig. 5 (b) changes with sampling rate with the Y-PSNR of the present invention, OMP method, IHT method and patented method reconstructed image respectively for Lena figure, the trend comparison diagram that wherein Fig. 5 (c) changes with sampling rate with the Y-PSNR of the present invention, OMP method, IHT method and patented method reconstructed image respectively for Pepper figure.

With the present invention, OMP method, IHT method and patented method are other, Barbara figure, Lena figure, Pepper figure are reconstructed, the Y-PSNR PSNR of the image of reconstruct is as shown in table 1 below:

The PSNR of table 1 the present invention and OMP method, IHT method, patented method reconstructed image under different sampling rates:

As can be seen from Table 1, under any sampling rate, the Y-PSNR PSNR of reconstructed image of the present invention is higher than OMP method, IHT method and patented method.Therefore,, no matter under which kind of sampling rate, the present invention has better reconstruct effect than the OMP method of prior art, IHT method and patented method on Y-PSNR PSNR.

(3) emulation 3

In this emulation, checking observation vector has local similarity, this character is used in the time using local growth method to carry out observation vector cluster, in addition the threshold value that should choose while having analyzed by experiment for observation vector cluster under different sampling rates, has verified that the present invention uses monatomic direction and polyatom direction and the necessity of the initialization of population scheme of depositing.Wherein, Fig. 6 (a1) is under 20% sampling rate, observation vector standard deviation difference statistics corresponding to all adjacent image pieces of test pattern Barbara, Fig. 6 (a2) is under 20% sampling rate, observation vector standard deviation difference statistics corresponding to all similar adjacent smooth image piece of test pattern Barbara, Fig. 6 (a3) is under 20% sampling rate, observation vector standard deviation difference statistics corresponding to all similar adjacent texture image piece of test pattern Barbara; Fig. 6 (b1) is under 30% sampling rate, observation vector standard deviation difference statistics corresponding to all adjacent image pieces of test pattern Barbara, Fig. 6 (b2) is under 30% sampling rate, observation vector standard deviation difference statistics corresponding to all similar adjacent smooth image piece of test pattern Barbara, Fig. 6 (b3) is under 30% sampling rate, observation vector standard deviation difference statistics corresponding to all similar adjacent texture image piece of test pattern Barbara.

From Fig. 6 (a1) and Fig. 6 (b1), find out, most standard deviation differences are less, and in numerical value 1, observation vector corresponding to this explanation adjacent image piece has local similarity, from Fig. 6 (a2) and Fig. 6 (a3) and Fig. 6 (b2) and Fig. 6 (b3), find out, similar adjacent smooth and standard deviation difference corresponding to texture block are also distributed between 0 to 1 mostly, therefore texture block is similar with the feature of smooth standard deviation difference, so cannot according to standard deviation difference judge observation vector corresponding be texture block or smooth, also just cannot be respectively according to the Orientation Features of smooth and the one direction feature initialization population of texture block, therefore the present invention is according to monatomic direction and polyatom direction the scheme initialization population of depositing, can take into account smooth and texture block, thereby obtain good reconstructed image.

In order to choose the threshold value in cluster, in Fig. 6 (a1), getting 25% of histogram top is dividing line, obtain standard deviation and be 0 to 0.5 all eligible, for threshold value further accurately, in Fig. 6 (a2), getting 80% of histogram top is dividing line, obtain standard deviation 0 to 0.3 all eligible, therefore, under 20% sampling rate, cluster threshold value chooses 0.3; In Fig. 6 (b1), getting 25% of histogram top is dividing line, obtain standard deviation and be 0 to 0.4 all eligible, for threshold value further accurately, in Fig. 6 (a2), getting 80% of histogram top is dividing line, obtains standard deviation 0 to 0.4 all eligible, therefore under 30% sampling rate, cluster threshold value chooses 0.4.

In sum, restructuring procedure of the present invention comprises original image piecemeal is observed; Utilize observation vector local similarity to use the method for local growth to carry out cluster, each class observation vector is used based on part and selects machine-processed optimized Genetic Algorithm to find out in direction preferably former sub-portfolio, preserve population; To each image block, carry out population extended operation, use clonal selection algorithm to find out optimum former sub-portfolio to each image block; Image block is reconstructed by the former sub-portfolio of optimum; All reconstructed image pieces are spelled to composition view picture reconstructed image.The present invention takes full advantage of direction character, the direction character of natural image and the part of local similarity and genetic algorithm of Ridgelet dictionary atom and selects mechanism, using the optimized algorithm in natural calculating field as nonlinear optimization reconstructing method, not only can be under low sampling rate the various natural images of reconstruct preferably, experiment showed, the reconstruct of various sampling rate hypographs is all had to the chromatic effect of going out.Solved the problem that image visual effect is poor, Y-PSNR is low that existing non-protruding compressed sensing reconstructing method reconstructs under low sampling rate, reconstruction accuracy of the present invention is high, and image visual effect is good, can be used for the non-protruding compressed sensing reconstruct of low sampling image signal.

Below the step of not describing in detail and specifically push over process and can find on online and open source information, so do not describe in detail.

Claims (5)

1. based on local similarity and a local non-protruding compressed sensing image reconstructing method of selecting, comprise the steps:

(1) take original image, and it is carried out observing and receiving after piecemeal;

(2) calculate the standard deviation of each observation vector and utilize the local similarity of observation vector, adopting the method for local growth to carry out cluster to the observation vector of all image blocks;

(3) to image block corresponding to each class observation vector by polyatom direction and monatomic direction the scheme initialization population of depositing;

(4) utilize improved evolutionary programming algorithm that the population in step (3) is intersected, made a variation and select machine-processed selection operation based on part, image block corresponding to the every class observation vector of reconstruct, obtains former sub-portfolio optimum in dictionary direction;

(5) recycling Immune Clone Selection optimized algorithm is learnt out former sub-portfolio optimum in yardstick and displacement;

(6) reconstructed image the piece corresponding all observation vectors that obtain in step 5 is stitched together according to the order of sequence and obtains view picture reconstructed image, output view picture reconstructed image.

2. non-protruding compressed sensing image reconstructing method according to claim 1, is characterized in that, the concrete steps of described step (2) are as follows:

(2.1) calculate the standard deviation of each observation vector;

(2.2) all image blocks are all arranged to a cluster mark, be initially all labeled as 0, wherein, mark 0 represents not by cluster, and mark 1 represents to be included in a certain classification;

(2.3), from first image block, successively each image block is done to following operation: if image block cluster is labeled as 1, do not operate, if be labeled as 0, carry out M _ithe cluster of class, clustering method is as follows:

(2.3.1) taking current image block i as seed image block, observation vector corresponding seed image block is added to class M _iand as Seeding vector;

(2.3.2) by the standard deviation of Seeding vector respectively with seed image block eight adjacent image piece A around ₁, A ₂a ₈the standard deviation of corresponding observation vector is subtracted each other and is obtained C ₁, C ₂c ₈; If image block A _i(i=1,2 ... 8) cluster be labeled as 0, and | C _i|≤τ (i=1,2 ... 8), wherein τ is threshold value, by image block A _icluster mark is set to 1, and by its corresponding observation vector y _iadd class M _iin; The observation vector that these are added is according to its respective standard difference | C _i| the order increasing progressively is at class M _imiddle arrangement;

(2.3.3) at class M _iin, if Seeding vector is last element, M _iclass cluster completes; Otherwise, make M _ifirst element after middle Seeding vector is new Seeding vector, and making image block corresponding to new seed vector is new seed image block, upgrades threshold tau=max (0.1, τ-0.1), then repeating step 2.3.2) and 2.3.3).

3. non-protruding compressed sensing image reconstructing method according to claim 1, is characterized in that, the related initialization of population scheme of described step (3) is as follows:

If the population of the image block that such observation vector is corresponding is A, in population, individual amount is (H+P); Again by following Optimization Steps:

(3.1) individual for front H, utilizing sliding window method is that each individuality is chosen 10 directions, and sliding window scope is 1 to P, and length of window is 10, and sliding window lap is 8; From corresponding each the sub-dictionary of these 10 directions, k atom of random generation forms this individuality;

(3.2) for below P individual, make each individuality comprise successively a direction in 1 to P, produce k atom and form this individuality at random from sub-dictionary corresponding to this direction.

4. non-protruding compressed sensing image reconstructing method according to claim 1, is characterized in that, the operation steps of described step (4) is as follows:

(4.1) calculate the fitness of all individualities in the population of the image block that each class observation vector is corresponding according to following fitness function:

Wherein, f (X _m) be m individual fitness value in the population A of the image block that such observation vector is corresponding, the label that i is observation vector, j is the sum of observation vector in each class after cluster, y _ifor i observation vector in class, Φ is Gauss's observing matrix, X _mm individual all gene position in population A, dec (X _m) representing these gene position corresponding former sub-portfolio in dictionary, this former sub-portfolio is exactly a sub-dictionary, α _ifor the rarefaction representation coefficient vector of image block corresponding to i observation vector in class, it is by the generalized inverse matrix of sensing matrix and this observation vector y _imultiply each other and obtain, sensing matrix is by Gauss's observing matrix Φ and sub-dictionary dec (X _m) multiplying each other obtains, be vectorial two norms square.

(4.2) population A of image block corresponding to such observation vector is carried out to the intersection of genetic algorithm, variation, selects operation, upgrades and preserves the population A after operating ₁;

(4.2.1) interlace operation: current population A is divided into two sub-populations, and first sub-population Q1 is made up of front H multi-direction individuality, second sub-population Q2 is made up of P one direction individuality below.For each individual i in sub-population Q1 prepares a set U _i, this set is initialized as that { i} is random one (0,1) the interval numerical value v that generates of individual i _iif, numerical value v _ibe less than crossover probability p _c, random another individuality of selection from sub-population Q1, and select at random a crossover location, these two individualities are carried out to single-point intersection and obtain two new individualities, add these two new individualities to set U _iin; In like manner, for each individual i in sub-population Q2 prepares a set V _i, this set is initialized as that { i} is random one (0,1) the interval numerical value v that generates of individual i _iif, numerical value v _ibe less than crossover probability p _c, random another individuality of selection from sub-population Q2, and select at random a crossover location, these two individualities are carried out to single-point intersection and obtain two new individualities, add these two new individualities to set V _iin;

(4.2.2) mutation operation: generate one (0,1) interval numerical value u for each individual i in sub-population Q1 is random _iif this numerical value is less than variation Probability p _m, to variation position of the random selection of this individuality, and select at random a numerical value that is not more than dictionary scale to replace the numerical value of variation position, obtain a new individuality, this individuality is added to the set U of its correspondence _iin; In like manner, generate one (0,1) interval numerical value u for each individual i in sub-population Q2 is random _iif this numerical value is less than variation Probability p _m, to variation position of the random selection of this individuality, and select at random a numerical value that is not more than dictionary scale to replace the numerical value of variation position, obtain a new individuality, this individuality is added to the set V of its correspondence _iin;

(4.2.3) local selection operates: be set U corresponding to the each individual i of sub-population Q1 _iin select the individuality of fitness minimum, with it replace the individual i in former population A; In like manner, be set V corresponding to the each individual i of sub-population Q2 _iin select the individuality of fitness minimum, replace the individual i in former population A with it, obtain new population A ₁.

(4.3) judge population A ₁whether meet the iteration stopping condition of genetic algorithm, if meet, go to step (4.4), if do not meet, by population A ₁as the population A of image block corresponding to such observation vector, go to step (4.1).

(4.4) individuality that selection fitness is the highest is as optimum individual, one group of corresponding with optimum individual the rarefaction representation coefficient vector of image block corresponding such each observation vector Ridgelet base atom is multiplied each other, obtain the reconstructed image piece that such observation vector is corresponding, preserve the population A after optimization renewal in step (4.2) ₂.

(4.5) population of image block corresponding to all class observation vectors is performed step to (4.1) successively to step (4.4), obtain the reconstructed image piece that all class observation vectors are corresponding and optimize the population after upgrading.

5. non-protruding compressed sensing image reconstructing method according to claim 1, is characterized in that, the operation steps of described step (5) is as follows:

(5.1) for the each reconstructed image piece B obtaining in step 4, by its corresponding population A of preserving in step (4.4) ₂as its initialization population;

(5.2) add two optimum antibody corresponding 8 adjacent image pieces of image block B to initialization population A ₂, and remove and repeat to obtain population A after antibody ₃, calculate population A ₃scale, be designated as l;

(5.3) first, calculate population A corresponding to this image block B according to following affinity function ₃in the affinity of all antibody:

Wherein, g (X _m) be population A corresponding to this image block B ₃in the affinity value of m antibody, y ₀be the observation vector of image block B, Φ is Gauss's observing matrix, X _mall gene position of m antibody in population A, dec (X _m) representing these gene position corresponding former sub-portfolio in dictionary, this former sub-portfolio is exactly a sub-dictionary, α _mfor image block B is at sub-dictionary dec (X _m) under rarefaction representation coefficient vector, it is by the generalized inverse matrix of sensing matrix and observation vector y ₀multiply each other and obtain, sensing matrix is by Gauss's observing matrix Φ and sub-dictionary dec (X _m) multiplying each other obtains, be vectorial two norms square;

(5.4) population A corresponding to image block B again ₃carry out the clone of clonal selection algorithm, variation, selects operation, upgrades and preserves the population A after operating ₄;

(5.5) judge whether to meet iteration stopping condition, if meet, go to step (5.6), if do not meet, by the population A of upgrading after operation ₄as population A corresponding to image block B ₃, go to step (5.3);

(5.6) antibody that selection affinity is the highest, as optimum antibody, multiplies each other one group of corresponding with optimum antibody the rarefaction representation coefficient vector of image block B Ridgelet base atom, obtains the reconstructed image piece of image block B, preserves the population A of optimizing after upgrading ₅;

(5.7) population of all observation vector correspondence image pieces is performed step to (5.1) successively to step (5.6), obtain the reconstructed image piece that all observation vectors are corresponding and optimize the population after upgrading.