CN106780500A - A kind of image partition method of use regression algorithm - Google Patents

Abstract

The present invention proposes a kind of image partition method of use regression algorithm, and its main contents includes：Background fitting smoothing model, pixel recurrence, RANSAC robust regressions, overall partitioning algorithm.Using recurrence partitioning algorithm come the difficulty of technology before overcoming, smooth treatment is carried out to background parts by fitting and smoothing model first, then pixel regression forecasting image pixel intensities are passed through, find the optimum value of model parameter, background is set not influenceed by foreground pixel, by RANSAC robust regression technologies, the number of exceptional value is reduced by the method for iteration to greatest extent, the speed of algorithm is improve by pretreatment, finally by overall partitioning algorithm, all pixels are used with least square fitting, image segmentation is completed.The algorithm is provided and has arrived excellent image segmentation performance, the difficulty of common methods before overcoming, and when background and foreground image scope have overlap, segmentation result also has preferably performance.

Description

A kind of image partition method of use regression algorithm

Technical field

The present invention relates to image processing field, more particularly, to a kind of image partition method of use regression algorithm.

Background technology

Image on display screen, such as computer and smart mobile phone, these images are generally mixing content file, generally comprise two Layer, is made up of the background and text, the prospect of bargraphs that smooth.Image segmentation algorithm is usually used in medical image segmentation, text The fields such as extraction, living things feature recognition and automatically texture segmentation, common image compression algorithm such as JPEG2000 and intraframe coding, But because prospect may be covered in the background of smooth change, color gamut and foreground color are overlapped so that segmentation is more difficult.

The present invention proposes a kind of image partition method of use regression algorithm, using recurrence partitioning algorithm come before overcoming Background parts are carried out smooth treatment by the difficulty of technology by fitting and smoothing model first, then by pixel regression forecasting picture Plain intensity, finds the optimum value of model parameter, background is not influenceed by foreground pixel, by RANSAC robust regression skills Art, the number of exceptional value is reduced by the method for iteration to greatest extent, and the speed of algorithm is improve by pretreatment, is finally led to Overall partitioning algorithm is crossed, all pixels are used with least square fitting, complete image segmentation.The algorithm is provided and has arrived excellent Image segmentation performance, the difficulty of common methods before overcoming, the segmentation result when background and foreground image scope have overlap There is preferably performance.

The content of the invention

To solve the above problems, the present invention provides a kind of image partition method of use regression algorithm, its main contents bag Include：

(1) Background fitting smoothing model；

(2) pixel is returned；

(3) RANSAC robust regressions；

(4) overall partitioning algorithm.

Wherein, described Background fitting smoothing model (), it will be assumed that, if an image block only has powerful connections, can be with As the basis of smooth model；The approximation of the pixel smooth function of each pixel has an error less than required threshold value； But if an image block is by the smooth background of some foreground pixels superposition, and these foreground pixels account for the ratio of block relatively Small, being now fitted smooth function will not represent these foreground pixels.

Further, each image is divided into nonoverlapping size block N × N and represents each image block, table by us F (x, y) is shown as, with smooth model B (x, y；α₁... α_K) treatment, use the linear combination P of some basic functions_k(x, y), then Using Karhunen-Loeve transformation training image, setting is optimized by smooth background；Using one group of two-dimensional dct as our smooth The linear combination of model, two-dimensional dct transform function is defined as：

P_{U, v (x, y)}=β_uβ_vcos((2x+1)πu/2N)cos((2y+1)πv/2N) (1)

Wherein u and v represent the frequency on basis, β_uAnd β_vIt is normalization factor, it is notable that based on supervision dictionary Practise the smooth expression for also contributing to produce background component with the algorithm of sub-space learning.

Wherein, described pixel is returned, and finds model parameter, it would be desirable to know which pixel belongs to background, find me Model parameter optimum value so that they not receive foreground pixel image.

Further, a solution of optimal model parameters is found, α is defined_k' s be cost function, it measures original The goodness of fit between the intensity of beginning pixel and the smooth model of prediction, then reduces cost function to greatest extent；One conjunction The cost function of reason can be L_p(P can be 0,1 to the error of fitting of standard, or 2), can be written as the solution：

F, α and P is set to represent the one-dimensional version of F, the size N of one matrix of vector sum of all parameters respectively²× K wherein K arranges corresponding P_kThe vector quantization version of (x, y), above mentioned problem can be expressed as：

α^*=argmin_α||f-Pα||_p (3)

If we use L now₂The cost function of standard (i.e. P=2), we have to asking to least square fitting Topic, and have the form solution of a closing as follows：

α=(P^TP)^-1P^Tf (4)

But be present a fact, i.e. model parameter in least square fitting, can be influenceed by foreground pixel, it is proposed that one Plant and be based on robustness regression method, reduce the quantity of model of fit exceptional value.

Wherein, described RANSAC robust regressions, RANSAC algorithms are a kind of sane regression algorithms, i.e., peeled off in presence Suitable model is found in one group of data of value；RANSAC algorithms are a kind of methods of iteration, different by reducing to greatest extent The number of constant value, carries out parameter Estimation；Partitioning algorithm in, can using foreground pixel as outlier smoothing model, RANSAC algorithms repeat to find one group of iterative program of data model.

Further, into nonoverlapping size block N × N, step is such as the foreground/background segmentation of the RANSAC algorithms of proposition Under：

1) a subset that K pixel is formed is randomly choosed, this subset is represented by following formula：

S={ (x_l, y_l), l=1,2 ..., K } (5)

2) by Matching ModelTo pixel (x_l, y_l) ∈ S, find α_k’s；By solving lower alignment Property equation：

∑_kα_kP_k(x_l, y_l)=F (x_l, y_l) (6)

L=1,2 ... K here F (x, y) represent pixel (x, y) brightness value；

3) N is tested²All of pixel F (x, y) to block model of fit, when prediction these pixels error less than overall Partitioning algorithm ∈_inWhen, now it is considered correct；

If its size is bigger than previous, the uniformity collection of current iteration is preserved；Repeat this proword when Wait, or when maximum common recognition accounts for whole data set certain proportion, labeled as ∈₂；After the completion of this process, set in maximum common recognition Pixel will be considered as selection or equivalence background.

Further, in the first step, the data fork according to a subset goes out the parameter of model, and in second step, test comes From the whole data set of subset, go to push up the quantity for seeing that simulation model is consistent；One sample is considered as an outlier, if it has One error of fitting is more than a threshold value, is defined as maximum allowable offset, and RANSAC repetitive routine fixed number of times is finally selected The Optimized model of the common recognition (interior point set) of model and maximum.

Wherein, described overall partitioning algorithm, it is characterised in that constant block, the background and text/graphics of smooth change are folded Constant background is added in, all of pixel of constant block has similar intensity；Standard deviation such as fruit block is less than certain threshold value, We announce the block for constant, and the background of smooth change is a block, and the Strength Changes of all pixels can be smooth by one Function model well, therefore, least square fitting is used to all pixels；If all pixels of the block can be with small In the misrepresentation ∈ of predetermined threshold_in, referred to as constant background, image block is for including text/graphics are superimposed upon constant background Generally there is zero variance (or very small variance), these images generally have limited amount in each block between portion's coupling assembly Different colors (typically smaller than 10), the intensity of different parts is very different；We calculate in the block each The percentage of individual different color, background will be selected as with highest percentage, and others are used as prospect, when one piece not Meet any of the above described condition, RANSAC will be applied onto in single background and prospect.

Further, the overall partitioning algorithm of each size block N × N is summarized as follows and (is only applicable to a coloured image Gray component algorithm)：

If 1) standard deviation of image pixel intensities is less than ∈₁, then whole block is marked as background；If it is not, carrying out down One step；

2) least square fitting is performed to all pixels, the error if all of pixel is both less than ∈_in, then by whole block It is marked as background；If it is not, carrying out next step；

If 3) number of different colours is higher than R less than T1 and strength range, text/graphics in a constant background and Color, then pixel ratio highest is background colour；If it is not, if it is not, carrying out next step；

4) background and prospect are split using RANSAC algorithms, pixel is less than ∈ with error of fitting_inBackground will be considered as.

Brief description of the drawings

Fig. 1 is a kind of system flow chart of the image partition method of use regression algorithm of the invention.

Fig. 2 is a kind of segmentation background layer result of the image partition method of use regression algorithm of the invention.

Fig. 3 is a kind of segmentation foreground layer result of the image partition method of use regression algorithm of the invention.

Specific embodiment

It should be noted that in the case where not conflicting, the feature in embodiment and embodiment in the application can phase Mutually combine, the present invention is described in further detail with specific embodiment below in conjunction with the accompanying drawings.

Fig. 1 is a kind of system flow chart of the image partition method of use regression algorithm of the invention.Main contents include：The back of the body Scape fitting and smoothing model, pixel recurrence, RANSAC robust regressions, overall partitioning algorithm.

Wherein, described Background fitting smoothing model (), it will be assumed that, if an image block only has powerful connections, can be with As the basis of smooth model；The approximation of the pixel smooth function of each pixel has an error less than required threshold value； But if an image block is by the smooth background of some foreground pixels superposition, and these foreground pixels account for the ratio of block relatively Small, being now fitted smooth function will not represent these foreground pixels.

Wherein, described pixel is returned, and finds model parameter, it would be desirable to know which pixel belongs to background, find me Model parameter optimum value so that they not receive foreground pixel image.

Further, a solution of optimal model parameters is found, α is defined_k' s be cost function, it measures original The goodness of fit between the intensity of beginning pixel and the smooth model of prediction, then reduces cost function to greatest extent；One conjunction The cost function of reason can be L_p(P can be 0,1 to the error of fitting of standard, or 2), can be written as the solution：

F, α and P is set to represent the one-dimensional version of F, the size N of one matrix of vector sum of all parameters respectively²× K wherein K arranges corresponding P_kThe vector quantization version of (x, y), above mentioned problem can be expressed as：

α^*=argmin_α||f-Pα||_p (3)

If we use L now₂The cost function of standard (i.e. P=2), we have to asking to least square fitting Topic, and have the form solution of a closing as follows：

α=(P^TP)^-1P^Tf (4)

But be present a fact, i.e. model parameter in least square fitting, can be influenceed by foreground pixel, it is proposed that one Plant and be based on robustness regression method, reduce the quantity of model of fit exceptional value.

Wherein, described RANSAC robust regressions, RANSAC algorithms are a kind of sane regression algorithms, i.e., peeled off in presence Suitable model is found in one group of data of value；RANSAC algorithms are a kind of methods of iteration, different by reducing to greatest extent The number of constant value, carries out parameter Estimation；Partitioning algorithm in, can using foreground pixel as outlier smoothing model, RANSAC algorithms repeat to find one group of iterative program of data model.

Further, into nonoverlapping size block N × N, step is such as the foreground/background segmentation of the RANSAC algorithms of proposition Under：

1) a subset that K pixel is formed is randomly choosed, this subset is represented by following formula：

S={ (x_l, y_l), l=1,2 ..., K } (5)

2) by Matching ModelTo pixel (x_l, y_l) ∈ S, find α_k’s；By solving lower alignment Property equation：

∑_kα_kP_k(x_l, y_l)=F (x_l, y_l) (6)

L=1,2 ... K here F (x, y) represent pixel (x, y) brightness value；

3) N is tested²All of pixel F (x, y) to block model of fit, when prediction these pixels error less than overall Partitioning algorithm ∈_inWhen, now it is considered correct；

If its size is bigger than previous, the uniformity collection of current iteration is preserved；Repeat this proword when Wait, or when maximum common recognition accounts for whole data set certain proportion, labeled as ∈₂；After the completion of this process, set in maximum common recognition Pixel will be considered as selection or equivalence background.

Further, in the first step, the data fork according to a subset goes out the parameter of model, and in second step, test comes From the whole data set of subset, go to push up the quantity for seeing that simulation model is consistent；One sample is considered as an outlier, if it has One error of fitting is more than a threshold value, is defined as maximum allowable offset, and RANSAC repetitive routine fixed number of times is finally selected The Optimized model of the common recognition (interior point set) of model and maximum.

Wherein, described overall partitioning algorithm, it is characterised in that constant block, the background and text/graphics of smooth change are folded Constant background is added in, all of pixel of constant block has similar intensity；Standard deviation such as fruit block is less than certain threshold value, We announce the block for constant, and the background of smooth change is a block, and the Strength Changes of all pixels can be smooth by one Function model well, therefore, least square fitting is used to all pixels；If all pixels of the block can be with small In the misrepresentation ∈ of predetermined threshold_in, referred to as constant background, image block is for including text/graphics are superimposed upon constant background Generally there is zero variance (or very small variance), these images generally have limited amount in each block between portion's coupling assembly Different colors (typically smaller than 10), the intensity of different parts is very different；We calculate in the block each The percentage of individual different color, background will be selected as with highest percentage, and others are used as prospect, when one piece not Meet any of the above described condition, RANSAC will be applied onto in single background and prospect.

Fig. 2 is a kind of segmentation background layer result of the image partition method of use regression algorithm of the invention.Use modulus of smoothness Type, figure is the background pixel detected using least square fitting.It can be seen that background layer is unusual light, there is no any sharp Edge.Each image is divided into nonoverlapping size block N × N and represents each image block by us, is expressed as F (x, y), With smooth model B (x, y；α₁... α_K) treatment, use the linear combination P of some basic functions_k(x, y), then using Karhunen-Loeve transformation Training image, setting is optimized by smooth background；Using one group of two-dimensional dct as linear group of our smoothing model Close, two-dimensional dct transform function is defined as：

P_{U, v (x, y)}=β_uβ_vcos((2x+1)πu/2N)cos((2y+1)πv/2N) (1)

Wherein u and v represent the frequency on basis, β_uAnd β_vIt is normalization factor, it is notable that based on supervision dictionary Practise the smooth expression for also contributing to produce background component with the algorithm of sub-space learning.

Fig. 3 is a kind of segmentation foreground layer result of the image partition method of use regression algorithm of the invention.As can be seen that In the case of all of, the algorithm provides superior performance, in the case of having challenge, when the background and foreground color scope of image When having overlap, the algorithm also has preferably performance.The overall partitioning algorithm of each size block N × N is summarized as follows and (is only applicable to One algorithm of the gray component of coloured image)：

If 1) standard deviation of image pixel intensities is less than ∈₁, then whole block is marked as background；If it is not, carrying out down One step；

2) least square fitting is performed to all pixels, the error if all of pixel is both less than ∈_in, then by whole block It is marked as background；If it is not, carrying out next step；

If 3) number of different colours is higher than R less than T1 and strength range, text/graphics in a constant background and Color, then pixel ratio highest is background colour；If it is not, if it is not, carrying out next step；

4) background and prospect are split using RANSAC algorithms, pixel is less than ∈ with error of fitting_inBackground will be considered as.

For those skilled in the art, the present invention is not restricted to the details of above-described embodiment, without departing substantially from essence of the invention In the case of god and scope, the present invention can be realized with other concrete forms.Additionally, those skilled in the art can be to this hair Bright to carry out various changes and modification without departing from the spirit and scope of the present invention, these improvement also should be regarded as of the invention with modification Protection domain.Therefore, appended claims are intended to be construed to include preferred embodiment and fall into all changes of the scope of the invention More and modification.

Claims (10)

1. a kind of image partition method of use regression algorithm, it is characterised in that Background fitting smoothing model ()；Pixel is returned (2)；RANSAC robust regressions (three)；Overall partitioning algorithm (four).

2. based on the Background fitting smoothing model () described in claims 1, it is characterised in that we assume that, if one Image block only has powerful connections, can be as the basis of smooth model；The approximation of the pixel smooth function of each pixel has one Error is less than required threshold value；But if an image block is by the smooth background of some foreground pixels superposition, and these prospects The ratio that pixel accounts for block is relatively small, and being now fitted smooth function will not represent these foreground pixels.

3. based on the smooth background described in claims 2, it is characterised in that specifically, be divided into for each image by we Nonoverlapping size block N × N simultaneously represents each image block, is expressed as F (x, y), with smooth model B (x, y；α₁... α_K) place Reason, uses the linear combination P of some basic functions_k(x, y), then using Karhunen-Loeve transformation training image, is entered by smooth background Row optimal design-aside；Using one group of two-dimensional dct as our smoothing model linear combination, two-dimensional dct transform function is defined as：

P_{U, v (x, y)}=β_uβ_vcos((2x+1)πu/2N)cos((2y+1)πv/2N) (1)

Wherein u and v represent the frequency on basis, β_uAnd β_vNormalization factor, it is notable that based on supervision dictionary learning and The algorithm of sub-space learning also contributes to produce the smooth expression of background component.

4. (two) are returned based on the pixel described in claims 1, it is characterised in that find model parameter, it would be desirable to know Which pixel belongs to background, finds the optimum value of our model parameter so that they do not receive foreground pixel image.

5. based on the optimal model parameters described in claims 4, a solution of optimal model parameters is found, define α_k’ S is cost function, and it measures the goodness of fit between the intensity of original pixels and the smooth model of prediction, then to greatest extent Reduce cost function in ground；One rational cost function can be L_p(P can be 0,1 to the error of fitting of standard, or 2), make this Solution can be written as：

$\{{\mathrm{\α}}_1^{*},\mathrm{...},{\mathrm{\α}}_K^{*}\}=\arg \underset{{\mathrm{\α}}_1,\mathrm{...},{\mathrm{\α}}_K}{\min }\underset{x,y}{\Σ}|F(x,y)-\underset{k=1}{\overset{K}{\Σ}}{\mathrm{\α}}_k{P}_k(x,y){|}^p---\left(2\right)$

F, α and P is set to represent the one-dimensional version of F, the size N of one matrix of vector sum of all parameters respectively²× K wherein kth row are right The P for answering_kThe vector quantization version of (x, y), above mentioned problem can be expressed as：

α^*=argmin_α||f-Pα||_p (3)

If we use L now₂The cost function of standard (i.e. P=2), we have to the problem of least square fitting, and have The form solution of one closing is as follows：

α=(P^TP)^-1P^Tf (4)

But be present a fact, i.e. model parameter in least square fitting, can be influenceed by foreground pixel, it is proposed that Yi Zhongji In robustness regression method, the quantity of model of fit exceptional value is reduced.

6. based on the RANSAC robust regressions (three) described in claims 1, it is characterised in that RANSAC algorithms are a kind of sane Regression algorithm, i.e., find suitable model in the one group of data that there is outlier；RANSAC algorithms are a kind of sides of iteration Method, by reducing the number of exceptional value to greatest extent, carries out parameter Estimation；In partitioning algorithm, foreground pixel can be made It is the smoothing model of outlier, RANSAC algorithms repeat to find one group of iterative program of data model.

7. based on the RANSAC algorithms described in claims 6, it is characterised in that the foreground/background of the RANSAC algorithms of proposition Nonoverlapping size block N × N is divided into, step is as follows：

1) a subset that K pixel is formed is randomly choosed, this subset is represented by following formula：

S={ (x_l, y_l), l=1,2 ..., K } (5)

2) by Matching ModelTo pixel (x_l, y_l) ∈ S, find α_k’s；By solving following linear side Journey：

∑_kα_kP_k(x_l, y_l)=F (x_l, y_l) (6)

L=1,2 ... K here F (x, y) represent pixel (x, y) brightness value；

3) N is tested²All of pixel F (x, y) to block model of fit, when the error of these pixels of prediction is split less than overall Algorithm ∈_inWhen, now it is considered correct；

If its size is bigger than previous, the uniformity collection of current iteration is preserved；When repeating this proword, or When maximum common recognition accounts for whole data set certain proportion, labeled as ∈₂；After the completion of this process, in the picture that maximum common recognition is set Element will be considered as selection or equivalence background.

8. based on the RANSAC algorithm steps described in claims 7, it is characterised in that in the first step, according to a subset Data fork goes out the parameter of model, in second step, tests the whole data set from subset, goes to push up the number for seeing that simulation model is consistent Amount；One sample is considered as an outlier, if it has an error of fitting more than a threshold value, is defined as maximum allowable inclined The Optimized model of the common recognition (interior point set) of difference, RANSAC repetitive routine fixed number of times, the model for finally selecting and maximum.

9. based on the overall partitioning algorithm (four) described in claims 1, it is characterised in that constant block, the background of smooth change Constant background is superimposed upon with text/graphics, all of pixel of constant block has similar intensity；Such as the standard deviation of fruit block Less than certain threshold value, we announce the block for constant, and the background of smooth change is a block, and the Strength Changes of all pixels can be with Modeled well by a smooth function, therefore, least square fitting is used to all pixels；If all pictures of the block Element can use the misrepresentation ∈ less than predetermined threshold_in, referred to as constant background, image block is superimposed upon perseverance for text/graphics Fixed background internally generally has zero variance (or very small variance) between coupling assembly, and these images are generally in each block In have a limited number of different colors (typically smaller than 10), the intensity of different parts is very different；We calculate The percentage of each the different color in the block, background will be selected as with highest percentage, and other conducts Prospect, any of the above described condition is unsatisfactory for when one piece, and RANSAC will be applied onto in single background and prospect.

10. based on the partitioning algorithm described in claims 9, it is characterised in that the overall segmentation of each size block N × N is calculated Method is summarized as follows (be only applicable to an algorithm for the gray component of coloured image)：

If 1) standard deviation of image pixel intensities is less than ∈₁, then whole block is marked as background；If it is not, carrying out next step；

2) least square fitting is performed to all pixels, the error if all of pixel is both less than ∈_in, then whole block is indicated It is background；If it is not, carrying out next step；

3) if the number of different colours is less than T1 and strength range is higher than R, text/graphics are in constant a background and face Color, then pixel ratio highest is background colour；If it is not, if it is not, carrying out next step；

4) background and prospect are split using RANSAC algorithms, pixel is less than ∈ with error of fitting_inBackground will be considered as.