CN104091305A - Quick image segmentation method used for computer graph and image processing and based on GPU platform and morphological component analysis - Google Patents

Abstract

The invention discloses a quick image segmentation method used for computer graph and image processing and based on a GPU platform and morphological component analysis. For a traditional image segmentation technology which is low in operation efficiency and poor in segmentation effect, the quick image segmentation method based on the GPU platform and the morphological component analysis is proposed so that operation time can be reduced effectively and the decomposition effect can be improved remarkably. The method includes the implementation steps that due to the fact that the advantages of GPU parallel operation are fully used, the algorithm of the morphological component analysis is achieved, the image segmentation efficiency is improved greatly, and thus quick image segmentation is achieved.

Description

A kind of for Fast image segmentation method computer graphic image processing, based on GPU platform and morphology component analysis

Technical field

The present invention relates to areas of information technology, further relate in digital image processing techniques field a kind of with Graphics Processing Unit (Graphics Processing Unit, GPU) realize the Fast Segmentation of image by form component analysis method (Morphological Component Analysis, MCA).This technology can be widely used in the fields such as image is cut apart, object detection and recognition.

Background technology

A different characteristic that major issue is differentiate between images in image processing and computer vision, the object that image is cut apart is the composition in separate picture with different characteristic, image can be made up of picture structure and image texture two parts, wherein, the geometric properties information that picture structure part has comprised image, by the region of Piecewise Smooth and clearly edge form.Image texture part is made up of high frequency oscillation component and the noise of image.In recent years, image is partitioned into a Disciplinary Frontiers in Low Level Vision and image processing, successfully image partition method has important using value in numerous Science and Technologies field, comprise pattern recognition system, Medical Image Processing, voice signal processing, communication system etc., become a common requirements of digital image processing field so improve the segmentation effect of image.

In recent years, the rarefaction representation method based on dictionary has obtained increasing application in image is processed.At dictionary design aspect, construct many kinds and effectively resolved dictionary, as wavelet field dictionary, discrete cosine dictionary, curve ripple dictionary etc., thereby propose based on rarefaction representation (SR, Sparse Representation) and the method cut apart of the image of morphology constituent analysis, the effect that image is cut apart improved greatly.But in the process of cutting apart due to MCA image, on the one hand, need two predefined dictionaries, and a texture part that is used for Description Image, another is used for the geometry part of Description Image.The effect of cutting apart in order to improve image, often the training of these dictionaries is made up of a large amount of training samples with similar content, and traditional dictionary training method need to consume a large amount of time.On the other hand, in each morphology component process of cutting apart image, image is extracted to each morphology component according to the atom in a given dictionary, then go to find the solution accepted of picture breakdown inverse problem according to sparse property constraint.In decomposable process, need said process to be carried out to the interative computation of hundreds of times, thereby consume a large amount of computing times.

Computer graphics processor (Graphics Processing Unit, GPU) high speed development, not only promoted the fast development of the applications such as image processing, virtual reality, Computer Simulation, the general-purpose computations of simultaneously also utilizing GPU to carry out beyond graphics process for people provides good operation platform.Graphics process based on GPU and general-purpose computations thereof become the hot research problem of graphics and high-performance computing sector.In recent years, along with the development of Graphics Processing Unit (GPU) and parallel processing technique, parallel image treatment technology arises at the historic moment, and become an important development direction of image processing field and computer science, gradually for the quick realization of various theories provides a brand-new and strong processing approach.GPU is extensively present in supercomputer at present, server, workstation, PC, the graphic process unit in mobile device even, the functions such as geometric transformation, illumination, triangular construction, cutting and drawing engine that it is integrated, and there is per second at least 1 thousand ten thousand polygonal processing power..GPU greatly promoted computer graphical processing speed, strengthened the quality of figure, and promoted and the fast development of relevant other application of computer graphical.With the serial design pattern difference of central processing unit (Central Processing Unit, CPU), GPU, for graphics process designs, has natural parallel computation characteristic.

Because view data itself is just huger, picture breakdown technology in addition is not often stinted the complexity that improves algorithm and is exchanged high-quality segmentation effect for.Because the time of processing is oversize, current existing image Segmentation Technology is that the serial arithmetic based on CPU realizes, and will spend a large amount of operation time.Not yet find at present on Patents or document the discussion for the Fast image segmentation method of GPU platform and morphology component analysis.

Summary of the invention

Goal of the invention: technical barrier to be solved by this invention be low for traditional image Segmentation Technology operation efficiency, segmentation effect is not good, propose a kind of Fast image segmentation method based on GPU platform and morphology component analysis, can effectively reduce operation time, improve significantly and decompose effect.

Realizing technical thought of the present invention is: utilize the parallel computation of GPU and the morphology constituent analysis of image to realize a kind of technology of Fast image segmentation, the time of Parallel Implementation was shortened greatly compared with the serial implementation time, thereby reach the object of Fast Segmentation image.Key step is as follows:

(1) carry out Memory Allocation optimization and initialization, image to be split is read in CPU internal memory, and initialization dictionary matrix;

According to resolution sizes unified distribution CPU and GPU end memory the initialization of input picture, and image to be split is read in CPU internal memory, and initialization dictionary matrix, Local Cosine Transform LDCT initialisation image texture part dictionary D used _t, use wavelet transformation initialisation image structure division dictionary D _n

(2) view data and initial dictionary are sent to GPU video memory from CPU internal memory;

By view data f to be split and initial dictionary D _t, D _nbe sent to GPU video memory from CPU internal memory.In image f, contain white Gaussian noise n, and image can be expressed as f=u+v+n, the structure division that wherein u is image, the texture part that v is image..

(3) the texture part v=v of still image on GPU platform ^(k), D _t, decomposite the structure division u=u of image ^(k+2);

Make D _n, D _tstructure division after representative study, the dictionary of texture part respectively, so order u _i, V _ibe illustrated respectively in the extraction block of pixels operator of picture structure part and i position of image texture part, λ ₁, λ ₂, λ ₃for balance parameters.The sub-optimization problem that solves picture structure part can be described as:

$\begin{array}{c}(u^{(k+1)},D_n,{\mathrm{\α}}_i)=\underset{u,D_n,{\mathrm{\α}}_i}{\arg \min }{\mathrm{\λ}}_1{\left|\right|(f-v^{\left(k\right)}-u)\left|\right|}^2+\underset{i}{\mathrm{\Σ}}{\left|\right|U_{i}u-D_n{\mathrm{\α}}_i\left|\right|}^2+\underset{i}{\mathrm{\Σ}}{u}_i{\left|{\mathrm{\α}}_i\right|}_0\\ s.t.d_l^T{d}_l=1,l=1,...,k_1\end{array}---\left(1\right)$

The dictionary D of this procedural learning to one presentation video structure division _n, and to picture structure part u ^(k)=f-v ^(k)do denoising, therefore obtain the structural images u of a denoising version ^(k+1), next remaining texture component separating in structural images is gone out, thereby obtains isolating the structural images u of texture composition ^(k+2).Sub-optimization problem can be described as:

$u^{(k+2)}=\underset{u}{\arg \min }({\mathrm{\λ}}_2{|\▿u|}_1+{\mathrm{\λ}}_3{{\left|\right|v\left|\right|}^2}_{H^{-1}}),s.t,v=u^{(k+1)}-u---\left(2\right)$

(4) the structure division u=u of still image on GPU platform ^(k+2), D _n, decomposite the texture part v=v of image ^(k+2);

The sub-optimization problem of this step can be described as:

$\begin{array}{c}(v^{(k+1)},D_t,{\mathrm{\β}}_i)=\underset{v,D_t,{\mathrm{\β}}_i}{\arg \min }{\mathrm{\λ}}_1{\left|\right|(f-u^{(k+2)}-v)\left|\right|}^2+\underset{i}{\mathrm{\Σ}}{\left|\right|V_{i}v-D_t{\mathrm{\β}}_i\left|\right|}^2+\underset{i}{\mathrm{\Σ}}{\mathrm{\η}}_i{\left|{\mathrm{\β}}_i\right|}_0,\\ s.t.s_m^T{s}_m=1,m=1,...,k_2\end{array}---\left(3\right)$

This sub-optimization problem is consistent with (1) formula, is divided into the reconstruction of texture dictionary learning and texture dictionary while solving.Next be that remaining constituent in texture image is separated, thereby guide the study of structure dictionary next time, obtain cutting apart rear texture image v=v simultaneously ^(k+2).The optimization problem that solves image texture part can be described as:

$v^{(k+2)}=\underset{v}{\arg \min }({\mathrm{\λ}}_2{|\▿u|}_1+{\mathrm{\λ}}_3{{\left|\right|v\left|\right|}^2}_{H^{-1}}),s.t,u=v^{(k+1)}-v---\left(4\right)$

Be back to step (3), after iteration (3), (4) two step k time, obtain picture structure part u and image texture part v after cutting apart.

(5) the structural images u after cutting apart and texture image v are sent back in the internal memory of CPU from the video memory of GPU, and show structural images and the texture image after cutting apart, discharge predefined internal memory simultaneously;

Compared with on going result, the invention has the beneficial effects as follows:

The present invention is cut apart owing to first morphology component analysis being applied to image, can effectively be partitioned into the structure division of image and the texture part of image;

The present invention is due to the framework that has adopted a kind of image of the GPU of utilization parallel computation to cut apart, and to existing the problems such as computing velocity is slow to realize the parallelization Fast Segmentation of image in serial algorithm, greatly shortened the working time that image is cut apart;

Brief description of the drawings

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is two test patterns to be split that use during the present invention tests;

Fig. 3 is that the present invention's the first width test pattern carries out structural images and the texture image after image is cut apart;

Fig. 4 is that the present invention's the second width test pattern carries out structural images and the texture image after image is cut apart;

Embodiment

(1) carry out Memory Allocation optimization and initialization, image to be split is read in CPU internal memory, and initialization dictionary matrix;

According to resolution sizes unified distribution CPU and GPU end memory the initialization of input picture, and image to be split is read in CPU internal memory, and initialization dictionary matrix, Local Cosine Transform LDCT initialisation image texture part dictionary D used _t, use wavelet transformation initialisation image structure division dictionary D _n

(2) view data and initial dictionary are sent to GPU video memory from CPU internal memory;

By view data f to be split and initial dictionary D _t, D _nbe sent to GPU video memory from CPU internal memory.In image f, contain white Gaussian noise n, and image can be expressed as f=u+v+n, the structure division that wherein u is image, the texture part that v is image..

(3) the texture part v=v of still image on GPU platform ^(k), D _t, decomposite the structure division u=u of image ^(k+2);

Make D _n, D _tstructure division after representative study, the dictionary of texture part respectively, so order u _i, V _ibe illustrated respectively in the extraction block of pixels operator of picture structure part and i position of image texture part, λ ₁, λ ₂, λ ₃for balance parameters.The sub-optimization problem that solves picture structure part can be described as:

$\begin{array}{c}(u^{(k+1)},D_n,{\mathrm{\α}}_i)=\underset{u,D_n,{\mathrm{\α}}_i}{\arg \min }{\mathrm{\λ}}_1{\left|\right|(f-v^{\left(k\right)}-u)\left|\right|}^2+\underset{i}{\mathrm{\Σ}}{\left|\right|U_{i}u-D_n{\mathrm{\α}}_i\left|\right|}^2+\underset{i}{\mathrm{\Σ}}{u}_i{\left|{\mathrm{\α}}_i\right|}_0\\ s.t.d_l^T{d}_l=1,l=1,...,k_1\end{array}---\left(1\right)$

The dictionary D of this procedural learning to one presentation video structure division _n, and to picture structure part u ^(k)=f-v ^(k)do denoising, therefore obtain the structural images u of a denoising version ^(k+1), next remaining texture component separating in structural images is gone out, thereby obtains isolating the structural images u of texture composition ^(k+2).Sub-optimization problem can be described as:

$u^{(k+2)}=\underset{u}{\arg \min }({\mathrm{\λ}}_2{|\▿u|}_1+{\mathrm{\λ}}_3{{\left|\right|v\left|\right|}^2}_{H^{-1}}),s.t,v=u^{(k+1)}-u---\left(2\right)$

(4) the structure division u=u of still image on GPU platform ^(k+2), D _n, decomposite the texture part v=v of image ^(k+2);

The sub-optimization problem of this step can be described as:

$\begin{array}{c}(v^{(k+1)},D_t,{\mathrm{\β}}_i)=\underset{v,D_t,{\mathrm{\β}}_i}{\arg \min }{\mathrm{\λ}}_1{\left|\right|(f-u^{(k+2)}-v)\left|\right|}^2+\underset{i}{\mathrm{\Σ}}{\left|\right|V_{i}v-D_t{\mathrm{\β}}_i\left|\right|}^2+\underset{i}{\mathrm{\Σ}}{\mathrm{\η}}_i{\left|{\mathrm{\β}}_i\right|}_0,\\ s.t.s_m^T{s}_m=1,m=1,...,k_2\end{array}---\left(3\right)$

This sub-optimization problem is consistent with (1) formula, is divided into the reconstruction of texture dictionary learning and texture dictionary while solving.Next be that remaining constituent in texture image is separated, thereby guide the study of structure dictionary next time, obtain cutting apart rear texture image v=v simultaneously ^(k+2).The optimization problem that solves image texture part can be described as:

$v^{(k+2)}=\underset{v}{\arg \min }({\mathrm{\λ}}_2{|\▿u|}_1+{\mathrm{\λ}}_3{{\left|\right|v\left|\right|}^2}_{H^{-1}}),s.t,u=v^{(k+1)}-v---\left(4\right)$

Be back to step (3), after iteration (3), (4) two step k time, obtain picture structure part u and image texture part v after cutting apart.

(5) the structural images u after cutting apart and texture image v are sent back in the internal memory of CPU from the video memory of GPU, and show structural images and the texture image after cutting apart, discharge predefined internal memory simultaneously;

The invention discloses a kind of for Fast image segmentation method computer graphic image processing, based on GPU platform and morphology component analysis.For more effective real-time computer graph and image processing, this invention has utilized the parallel computing of GPU to realize the Fast Segmentation of image, thereby has improved the efficiency of cutting apart of image.

Claims (2)

1. for Fast image segmentation method computer graphic image processing, based on GPU platform and morphology component analysis, it is characterized in that, comprise the steps:

1) carry out Memory Allocation optimization and initialization, image to be split is read in CPU internal memory, and initialization dictionary matrix;

According to resolution sizes unified distribution CPU and GPU end memory the initialization of input picture, to program end of run, do not carrying out Memory Allocation ever since; And image to be decomposed is read in CPU internal memory, and initialization dictionary matrix, use Local Cosine Transform LDCT initialisation image texture part dictionary D _t, use wavelet transformation initialisation image structure division dictionary D _n;

2) view data and initial dictionary are sent to GPU video memory from CPU internal memory;

By view data f to be split and initial dictionary D _t, D _nbe sent to GPU video memory from CPU internal memory; Suppose and in image f, contain white Gaussian noise n, and image can be expressed as f=u+v+n, the structure division that wherein u is image, the texture part that v is image; .

3) the texture part v=v of still image on GPU platform ^(k), D _t, decomposite the structure division u=u of image ^(k+2);

Make D _n, D _tstructure division after representative study, the dictionary of texture part respectively, so order u _i, V _ibe illustrated respectively in the extraction block of pixels operator of structure division and i position of texture part, λ ₁, λ ₂, λ ₃for balance parameters.The sub-optimization problem that solves picture structure part can be described as:

This procedural learning to one represents the dictionary D of structure division _n, and to initial configuration image u ^(k)=f-v ^(k)do denoising, therefore obtain the structural images u of a denoising version ^(k+1), next remaining texture component separating in structural images is gone out, thereby obtains isolating the structural images u of texture composition ^(k+2); Sub-optimization problem can be described as:

4) the structure division u=u of still image on GPU platform ^(k+2), D _n, decomposite the texture part v=v of image ^(k+2);

The sub-optimization problem of this step can be described as:

This sub-optimization problem is consistent with (1) formula, is divided into the reconstruction of texture dictionary learning and texture dictionary while solving; Next be that remaining constituent in texture image is separated, thereby guide the study of structure dictionary next time, obtain cutting apart rear texture image v=v simultaneously ^(k+2); The optimization problem that solves image texture part can be described as:

Be back to step (3), after iteration (3), (4) two step k time, obtain picture structure part u and image texture part v after cutting apart;

5) the structural images u after cutting apart and texture image v are sent back in the internal memory of CPU from the video memory of GPU, and show structural images and the texture image after cutting apart, discharge predefined internal memory simultaneously.

2. according to claim 1 a kind of for Fast image segmentation method computer graphic image processing, based on GPU platform and morphology component analysis, it is characterized in that: on GPU platform, realized morphology component analysis, the texture part of still image, decomposite the structure division of image and on GPU platform the structure division of still image, decomposite the methods such as the texture part of image.