CN105205485A - Large scale image segmentation algorithm based on multi-inter-class maximum variance algorithm - Google Patents

Abstract

The invention discloses a large scale image segmentation algorithm based on a multi-inter-class maximum variance algorithm. The multi-inter-class maximum variance algorithm is adopted to analyze a gray level image, an interesting gray level image area is obtained and divided into non-overlapped subblocks, an area growth method is adopted for segmenting the current gray level image, all pixels are comprehensively considered, the pixel segmentation accuracy is improved, and defects of an existing segmentation method are overcome. The large scale image segmentation algorithm can effectively segment a large scale image, is obvious in segmentation effect and can improve image segmentation precision.

Description

Based on the large scale image partitioning algorithm of maximum variance algorithm between multiclass class

Technical field

The invention belongs to image processing field, be specifically related to a kind of large scale image partitioning algorithm based on maximum variance algorithm between multiclass class.

Background technology

Iamge Segmentation refers to and utilizes some feature in image information, extracts user's interesting target from image.Information analysis is carried out to image to be split, feature extraction is carried out to important information.Before carrying out initialize partition, first to pixel extraction feature each in image, after feature extraction, can obtain a corresponding characteristic image, this characteristic image comprises three passages, and each passage correspond to a stack features value.Such as, but to the image of different scene, its characteristic image has different characteristics, has the image of strong texture features, its direction character can have larger contrast compared with other two features; And for natural scene image, the contrast in its color and direction can be larger.Put it briefly, certain consistance attribute criterion (tolerance) P given, is correctly divided into mutually the set of regions { S of not crossover by image X ₁, S ₂..., S _nprocess be referred to as segmentation, correct segmentation should meet following condition:

(3)P(S _i)＝1(true)

(4)P(S _i∪S _j)＝0(false)

Measurement criterion as Iamge Segmentation is not unique, it is relevant with application scenarios image and application purpose, and the scene image characteristic information for Iamge Segmentation has brightness, color, texture, structure, temperature, frequency spectrum, motion, shape, position, gradient and model etc.

The basic thought of maximum variance process image between class is utilized to be that the secondary gray level image of hypothesis one is made up of bright, dark two parts, if need two parts region segmentation to come, then the statistic histogram of the gray-scale value of entire image is divided into two parts, asks this two-part variance and obtain optimal threshold.

Between two class classes, maximum variance is described below: set gray-scale map gray level as L, carry out statistics with histogram, then it is divided into ξ according to threshold value t to it ₁, ξ ₂two classes, its probability occurred respectively is shown below:

$P_{{\mathrm{\ξ}}_1}={\mathrm{\Σ}}_{i=0}^t{p}_i,P_{{\mathrm{\ξ}}_2}={\mathrm{\Σ}}_{i=t+1}^{L-1}{p}_i=1-P_{{\mathrm{\ξ}}_1}$

ξ ₁, ξ ₂the average of two classes is respectively , can obtain:

${\mathrm{\μ}}_{{\mathrm{\ξ}}_1}={\mathrm{\Σ}}_{i=0}^t{\mathrm{ip}}_i/P_{{\mathrm{\ξ}}_1},{\mathrm{\μ}}_{{\mathrm{\ξ}}_2}={\mathrm{\Σ}}_{i=t+1}^{L-1}{\mathrm{ip}}_i{P}_{{\mathrm{\ξ}}_2}$

Therefore, global image pixel average is:

${\mathrm{\μ}}_0={\mathrm{\Σ}}_{i=0}^{L-1}{\mathrm{ip}}_i$

ξ ₁, ξ ₂variance between two classes is:

${\mathrm{\σ}}_{{\mathrm{\ξ}}_1,{\mathrm{\ξ}}_2}^2=P_{{\mathrm{\ξ}}_1}{({\mathrm{\μ}}_{{\mathrm{\ξ}}_1}-{\mathrm{\μ}}_0)}^2+P_{{\mathrm{\ξ}}_2}{({\mathrm{\μ}}_{{\mathrm{\ξ}}_2}-{\mathrm{\μ}}_0)}^2$

Between class, maximum variance can do multiclass and promotes, but asks during maximum variance multiclass and have two key issues, and how the first determines the number of classifying, and it two is that multiclass asks the complexity of maximum variance larger.By the restriction of above-mentioned two key issues, the existing partitioning algorithm of maximum variance process image between class that utilizes has been difficult to extract the appropriateness of image key feature information, and its partitioning algorithm is complicated simultaneously, and cause the resolution of the image of segmentation low, sharpness is not high.

Summary of the invention

Technical matters to be solved by this invention is: for the deficiencies in the prior art, provides a kind of large scale image partitioning algorithm based between multiclass class maximum variance algorithm high to large scale image segmentation precision.

The present invention solves the problems of the technologies described above adopted technical scheme: based on the large scale image partitioning algorithm of maximum variance algorithm between multiclass class, comprise the following steps:

(1) obtain the large scale scan image to be split that a width derives from high-definition camera or CCD camera, then convert this large scale scan image to be split to gray level image, then adopt maximum variance algorithm between multiclass class to analyze gray level image, obtain initial interested profile { A1, A2, A3 ... An}, wherein A1, A2, A3,, An is defined as all point forming initial interested profile;

(2) according to initial interested profile { A1, A2, A3,, An} obtains the area-of-interest in gray level image through envelope, choose a square region that this area-of-interest can be included according to the profile of this area-of-interest, the width defining this square region is W, is highly H, if W × H can be divided exactly by u × u, then this square region is defined as current gray level image, then direct current gray level image is divided into the size of individual non-overlapping copies is the sub-block of u × u; If W × H can not be divided exactly by u × u, then expand this square region and its size is divided exactly by u × u, the square region after expansion is defined as current gray level image, then current gray level image is divided into the size of individual non-overlapping copies is the sub-block of u × u, wherein, W' and H' correspondence represents width and the height of the square region after expansion, W'==W and H'>H or W'>W and H'==H or W'>W and H'>H, u get 4,5 or 6;

(3) adopt region-growing method to current gray level Image Segmentation Using, obtain the multiple preliminary aim regions in current gray level image;

(4) pending preliminary aim region current in current gray level image is defined as current preliminary target area;

(5) from current gray level image, extracting all sizes corresponding with current preliminary target area is the sub-block of u × u, the sub-block being u × u by each size that ordered pair current preliminary target area is corresponding processes, and the sub-block being u × u by current pending size is defined as current sub-block;

(6) be optimized in input parameter unbalanced input Optimized model by the pixel value separately of all pixels in current sub-block, this Non-linear Optimal Model is:

$\frac{ds}{dt}=A\×sin(2\mathrm{\π}\×f_0\×t+\mathrm{\ψ})+m\×g\left(t\right)+n\×s-m\×s^3+\sqrt{2\mathrm{\α}}\×\mathrm{\ξ}\left(t\right)$

Wherein: represent the signal to noise ratio (S/N ratio) that Non-linear Optimal Model exports, A is the amplitude of fixed cycle signal, f ₀for the frequency of fixed cycle signal, t is the run duration of Brownian Particles, ψ is the initial phase of fixed cycle signal, m, n are bistable state potential barrier real parameter, g (t) represents the input parameter of Non-linear Optimal Model, s is the coordinates of motion of Brownian Particles, and α is noise intensity, ξ (t) for average be the white Gaussian noise of 0;

After the pixel value separately of all pixels in current sub-block is optimized in input parameter unbalanced input Optimized model, Non-linear Optimal Model exports all pixels signal to noise ratio (S/N ratio) separately in current sub-block, if the signal to noise ratio (S/N ratio) of each pixel in current sub-block and to be stored in the large scale scan image in database in advance gray level image in the error of signal to noise ratio (S/N ratio) of corresponding pixel points be less than 10%, then determine current sub-block optimization success, wherein, the size being stored in the gray level image of the large scale scan image in database is in advance identical with the size of current gray level image,

(7) be that the sub-block of u × u is as current sub-block using size pending for the next one, then return step (6) to continue to perform, until the sub-block that all sizes corresponding to current preliminary target area are u × u is disposed, obtain corresponding final goal region;

(8) using pending preliminary aim region next in current gray level image as current preliminary target area, then return step (5) to continue to perform, until all preliminary aim regional processings in current gray level image are complete, obtain the multiple final goal regions in current gray level image, so far complete the segmentation of large scale image.

As preferably, the concrete acquisition process being stored in the signal to noise ratio (S/N ratio) of each pixel in the gray level image of the large scale scan image in database in step (6) is in advance:

(6.1) choose the representative large scale scan image that a width derives from high-definition camera or CCD camera, then this large scale scan image is converted to gray level image;

(6.2) width defining this gray level image is W, is highly H, if W × H can be divided exactly by u × u, then using this gray level image as pending gray level image, then direct pending gray level image to be divided into the size of individual non-overlapping copies is the sub-block of u × u; If W × H can not be divided exactly by u × u, then expand this gray level image and its size is divided exactly by u × u, using the gray level image after expansion as pending gray level image, then pending gray level image is divided into the size of individual non-overlapping copies is the sub-block of u × u, wherein, W' and H' correspondence represents width and the height of the gray level image after expansion, W'==W and H'>H or W'>W and H'==H or W'>W and H'>HW'>W and H'>H, u get 4,5 or 6;

(6.3) each target area in pending gray level image is manually drawn a circle to approve;

(6.4) signal to noise ratio (S/N ratio) that each size corresponding to each target area in pending gray level image is each pixel in the sub-block of u × u is calculated.

Compared with prior art, the invention has the advantages that: the large scale image partitioning algorithm based on maximum variance algorithm between multiclass class disclosed by the invention, maximum variance algorithm between multiclass class is adopted to analyze gray level image, obtain interested gray level image region, and be the sub-block of non-overlapping copies by interested gray level image Region dividing, adopt region-growing method to current gray level Image Segmentation Using, all pixels are considered, improve the accuracy of pixel segmentation, avoid the deficiency of existing dividing method; Partitioning algorithm of the present invention effectively can split large scale image, and segmentation effect is obvious, can improve Iamge Segmentation precision.

Accompanying drawing explanation

Fig. 1 is the large scale image of the embodiment before partitioning algorithm segmentation of the present invention;

Fig. 2 is the large scale image of the embodiment after partitioning algorithm segmentation of the present invention.

Embodiment

Below in conjunction with accompanying drawing embodiment, the present invention is described in further detail.

For the large scale image shown in Fig. 1, be the gray-scale map of the real road image of a width high-definition camera shooting, adopt the large scale image partitioning algorithm based on maximum variance algorithm between multiclass class of the present invention, comprise the following steps:

(1) obtain the scan image of Fig. 1, then convert this scan image to gray level image, then adopt maximum variance algorithm between multiclass class to analyze gray level image, obtain initial interested profile { A1, A2, A3 ... An}, wherein A1, A2, A3,, An is defined as all point forming initial interested profile;

(2) according to initial interested profile { A1, A2, A3,, An} obtains the area-of-interest in gray level image through envelope, choose a square region that this area-of-interest can be included according to the profile of this area-of-interest, the width defining this square region is W, is highly H, if W × H can be divided exactly by 5 × 5, then this square region is defined as current gray level image, then direct current gray level image is divided into the size of individual non-overlapping copies is the sub-block of 5 × 5; If W × H can not be divided exactly by 5 × 5, then expand this square region and its size is divided exactly by 5 × 5, the square region after expansion is defined as current gray level image, then current gray level image is divided into the size of individual non-overlapping copies is the sub-block of 5 × 5, wherein, W' and H' correspondence represents width and the height of the square region after expansion, W'==W and H'>H or W'>W and H'==H or W'>W and H'>H;

(3) adopt region-growing method to current gray level Image Segmentation Using, obtain the multiple preliminary aim regions in current gray level image;

(4) pending preliminary aim region current in current gray level image is defined as current preliminary target area;

(5) from current gray level image, extracting all sizes corresponding with current preliminary target area is the sub-block of 5 × 5, the sub-block being 5 × 5 by each size that ordered pair current preliminary target area is corresponding processes, and the sub-block being 5 × 5 by current pending size is defined as current sub-block;

(6) be optimized in input parameter unbalanced input Optimized model by the pixel value separately of all pixels in current sub-block, this Non-linear Optimal Model is:

$\frac{ds}{dt}=A\×sin(2\mathrm{\π}\×f_0\×t+\mathrm{\ψ})+m\×g\left(t\right)+n\×s-m\×s^3+\sqrt{2\mathrm{\α}}\×\mathrm{\ξ}\left(t\right)$

Wherein: represent the signal to noise ratio (S/N ratio) that Non-linear Optimal Model exports, A is the amplitude of fixed cycle signal, f ₀for the frequency of fixed cycle signal, t is the run duration of Brownian Particles, ψ is the initial phase of fixed cycle signal, m, n are bistable state potential barrier real parameter, g (t) represents the input parameter of Non-linear Optimal Model, s is the coordinates of motion of Brownian Particles, and α is noise intensity, ξ (t) for average be the white Gaussian noise of 0;

After the pixel value separately of all pixels in current sub-block is optimized in input parameter unbalanced input Optimized model, Non-linear Optimal Model exports all pixels signal to noise ratio (S/N ratio) separately in current sub-block, if the signal to noise ratio (S/N ratio) of each pixel in current sub-block and to be stored in the large scale scan image in database in advance gray level image in the error of signal to noise ratio (S/N ratio) of corresponding pixel points be less than 10%, then determine current sub-block optimization success, wherein, the size being stored in the gray level image of the large scale scan image in database is in advance identical with the size of current gray level image,

(7) be that the sub-block of 5 × 5 is as current sub-block using size pending for the next one, then return step (6) to continue to perform, until the sub-block that all sizes corresponding to current preliminary target area are 5 × 5 is disposed, obtain corresponding final goal region;

(8) using pending preliminary aim region next in current gray level image as current preliminary target area, then return step (5) to continue to perform, until all preliminary aim regional processings in current gray level image are complete, obtain the multiple final goal regions in current gray level image, so far complete the segmentation of the large scale image shown in Fig. 1.

In the step (6) of above embodiment, the concrete acquisition process being stored in the signal to noise ratio (S/N ratio) of each pixel in the gray level image of the large scale scan image in database is in advance:

(6.1) choose the scan image that a width derives from the representative road image of high-definition camera, then scan image is converted to gray level image;

(6.2) width defining this gray level image is W, is highly H, if W × H can be divided exactly by 5 × 5, then using this gray level image as pending gray level image, then direct pending gray level image to be divided into the size of individual non-overlapping copies is the sub-block of 5 × 5; If W × H can not be divided exactly by 5 × 5, then expand this gray level image and its size is divided exactly by 5 × 5, using the gray level image after expansion as pending gray level image, then pending gray level image is divided into the size of individual non-overlapping copies is the sub-block of 5 × 5, wherein, W' and H' correspondence represents width and the height of the gray level image after expansion, W'==W and H'>H or W'>W and H'==H or W'>W and H'>H;

(6.3) each target area in pending gray level image is manually drawn a circle to approve;

(6.4) signal to noise ratio (S/N ratio) that each size corresponding to each target area in pending gray level image is each pixel in the sub-block of 5 × 5 is calculated.

In partitioning algorithm of the present invention, between the multiclass class used, the method for maximum variance algorithm, region-growing method and expansion gray level image all adopts prior art.Such as, between multiclass class, maximum variance algorithm can with reference to " the chalk rice grain detection method based on multimodal distribution maximum between-cluster variance ", Xu Jiandong etc., optoelectronic laser, the 23rd volume the 5th phase, in May, 2012.

For Fig. 1, the design parameter of Non-linear Optimal Model is chosen as: keep amplitude A=4 of fixed cycle signal, the frequency f of fixed cycle signal ₀initial phase ψ=0 of=1Hz, fixed cycle signal is constant, and make noise intensity α span be [0,400], this seasonal bistable state potential barrier real parameter n=1, and make bistable state potential barrier real parameter m carry out the change that stepping is 0.1 within [1,10], supervisory system output signal-to-noise ratio simultaneously, when output signal-to-noise ratio curve produces characteristic peak and peak value is maximal value, namely can determine m=6.5, now parameters is optimization selection.

Utilize partitioning algorithm of the present invention and split Fig. 1 in conjunction with above-mentioned design parameter, the image after segmentation is shown in Fig. 2, as apparent from Fig. 2 can, after algorithm of the present invention segmentation, in Fig. 1, ground road sign is effectively split from road surface, and segmentation precision is high.

Claims (2)

1., based on the large scale image partitioning algorithm of maximum variance algorithm between multiclass class, it is characterized in that, comprise the following steps:

(1) obtain the large scale scan image to be split that a width derives from high-definition camera or CCD camera, then convert this large scale scan image to be split to gray level image, then adopt maximum variance algorithm between multiclass class to analyze gray level image, obtain initial interested profile { A1, A2, A3 ... An}, wherein A1, A2, A3,, An is defined as all point forming initial interested profile;

(2) according to initial interested profile { A1, A2, A3,, An} obtains the area-of-interest in gray level image through envelope, choose a square region that this area-of-interest can be included according to the profile of this area-of-interest, the width defining this square region is W, is highly H, if W × H can be divided exactly by u × u, then this square region is defined as current gray level image, then direct current gray level image is divided into the size of individual non-overlapping copies is the sub-block of u × u; If W × H can not be divided exactly by u × u, then expand this square region and its size is divided exactly by u × u, the square region after expansion is defined as current gray level image, then current gray level image is divided into the size of individual non-overlapping copies is the sub-block of u × u, wherein, W' and H' correspondence represents width and the height of the square region after expansion, W'==W and H'>H or W'>W and H'==H or W'>W and H'>H, u get 4,5 or 6;

(3) adopt region-growing method to current gray level Image Segmentation Using, obtain the multiple preliminary aim regions in current gray level image;

(4) pending preliminary aim region current in current gray level image is defined as current preliminary target area;

(5) from current gray level image, extracting all sizes corresponding with current preliminary target area is the sub-block of u × u, the sub-block being u × u by each size that ordered pair current preliminary target area is corresponding processes, and the sub-block being u × u by current pending size is defined as current sub-block;

(6) be optimized in input parameter unbalanced input Optimized model by the pixel value separately of all pixels in current sub-block, this Non-linear Optimal Model is:

$\frac{ds}{dt}=A\×sin(2\mathrm{\π}\×f_0\×t+\mathrm{\ψ})+m\×g\left(t\right)+n\×s-m\×s^3+\sqrt{2\mathrm{\α}}\×\mathrm{\ξ}\left(t\right)$

Wherein: represent the signal to noise ratio (S/N ratio) that Non-linear Optimal Model exports, A is the amplitude of fixed cycle signal, f ₀for the frequency of fixed cycle signal, t is the run duration of Brownian Particles, ψ is the initial phase of fixed cycle signal, m, n are bistable state potential barrier real parameter, g (t) represents the input parameter of Non-linear Optimal Model, s is the coordinates of motion of Brownian Particles, and α is noise intensity, ξ (t) for average be the white Gaussian noise of 0;

After the pixel value separately of all pixels in current sub-block is optimized in input parameter unbalanced input Optimized model, Non-linear Optimal Model exports all pixels signal to noise ratio (S/N ratio) separately in current sub-block, if the signal to noise ratio (S/N ratio) of each pixel in current sub-block and to be stored in the large scale scan image in database in advance gray level image in the error of signal to noise ratio (S/N ratio) of corresponding pixel points be less than 10%, then determine current sub-block optimization success, wherein, the size being stored in the gray level image of the large scale scan image in database is in advance identical with the size of current gray level image,

(7) be that the sub-block of u × u is as current sub-block using size pending for the next one, then return step (6) to continue to perform, until the sub-block that all sizes corresponding to current preliminary target area are u × u is disposed, obtain corresponding final goal region;

(8) using pending preliminary aim region next in current gray level image as current preliminary target area, then return step (5) to continue to perform, until all preliminary aim regional processings in current gray level image are complete, obtain the multiple final goal regions in current gray level image, so far complete the segmentation of large scale image.

2. the large scale image partitioning algorithm based on maximum variance algorithm between multiclass class according to claim 1, is characterized in that: the concrete acquisition process being stored in the signal to noise ratio (S/N ratio) of each pixel in the gray level image of the large scale scan image in database in step (6) is in advance:

(6.1) choose the representative large scale scan image that a width derives from high-definition camera or CCD camera, then this large scale scan image is converted to gray level image;

(6.2) width defining this gray level image is W, is highly H, if W × H can be divided exactly by u × u, then using this gray level image as pending gray level image, then direct pending gray level image to be divided into the size of individual non-overlapping copies is the sub-block of u × u; If W × H can not be divided exactly by u × u, then expand this gray level image and its size is divided exactly by u × u, using the gray level image after expansion as pending gray level image, then pending gray level image is divided into the size of individual non-overlapping copies is the sub-block of u × u, wherein, W' and H' correspondence represents width and the height of the gray level image after expansion, W'==W and H'>H or W'>W and H'==H or W'>W and H'>HW'>W and H'>H, u get 4,5 or 6;

(6.3) each target area in pending gray level image is manually drawn a circle to approve;

(6.4) signal to noise ratio (S/N ratio) that each size corresponding to each target area in pending gray level image is each pixel in the sub-block of u × u is calculated.