CN104484665A - Image foreground extracting method based on Gaussian variation model - Google Patents

Abstract

The invention discloses a novel method for automatic foreground extraction. Under the general conditions, the foreground extraction is interactive, and a foreground region needs an interactive labeling process. According to the method, an image is subjected to Gaussian variation processing to obtain a Gaussian variation point, then, the Gaussian variation point is subjected to marginalization processing by combining a Ncut technology, and marginal information of the image is obtained; finally, the center focus of the image is selected by a certain method, in addition, a key point is subjected to weighted growth in a direction towards the center focus, and a final foreground positioning point is obtained. The foreground region of a target image can be determined according to the proportion of the foreground positioning point in different regions. The whole process is fully automatic, and any manual interaction is not needed. Lots of experiments prove that the method is very effective, and in addition, more precise experiment results can be obtained.

Description

A kind of display foreground extracting method of the model that is deteriorated based on Gauss

Technical field

The present invention relates generally to image processing techniques, is specifically related to a kind of foreground extracting method be deteriorated based on Gauss.

Background technology

In the research of Computer Image Processing, the task of foreground extraction is by separated for interested object and background from single image or sequence image, for follow-up process.This is a flourishing long time study on classics problem, is also one of most basic research topic of Computer Image Processing.Interesting target generally refers to the moving object in the visual field, needs by analytical sequence Image Acquisition; Or the object of particular type, as face, vehicle etc., can by analyzing single width or sequence image extraction.Effective extraction of foreground area is extremely important for the task that target classification, identification and behavior understanding etc. are on the middle and senior level, because follow-up processing procedure only considers the pixel corresponding to foreground area in image usually, and the research of Video coding, retrieval, man-machine interaction, the problem such as motion-captured will be promoted greatly.But, owing to often there is many uncontrollable factors in actual acquisition environment, as the impact of the problems such as camera motion, illumination variation, shade, complex background, make Objective extraction fast and accurately become a difficult job.

Current foreground extracting method roughly can be divided into two classes: one is interactively foreground extraction, and another kind is automatic foreground extraction.Although these two kinds of methods are all for foreground extraction, the emphasis of these two kinds of methods is not identical.Interactively foreground extracting method lays particular emphasis on and takes foreground area exactly, does accurate process to the edge details of prospect.Interactively method needs artificial marking in target area or fringe region.But for the data of magnanimity on current internet, we obviously cannot carry out artificial mark, so we go to explore automatic foreground extraction.Automatic foreground extraction lays particular emphasis on the foreground area in positioning image, does not consider its details, and this process in interactively method by manually having marked.Automatic foreground extraction has a very wide range of applications prospect, such as image enhaucament, target identification, image index, the aspects such as content-based retrieval.

In these class methods of automatic foreground extraction, many experts have been had to explore.2011, Sungheum proposed a kind of method automatically extracting foreground object from multiple viewpoint.2012, the people such as Zhang Yupeng proposed the automatic prospect background cutting techniques of the degree of depth based on coded aperture.2013, the people such as Xie Chang-ting proposed an automatic trimap generation technique.Although said method achieves good effect, the accuracy and precision identified all has much room for improvement.

Summary of the invention

The object of the invention is to propose a kind of foreground extracting method be deteriorated based on Gauss, to improve the accuracy and precision of extraction prospect.In the invention, this implementation procedure is all completed automatically by computing machine, and user only needs to input target image, just can allow computer automatic analysis, final acquisition display foreground region.

Technical scheme of the present invention is as follows:

Step 1, uses Ncut technology to carry out region segmentation to target image, obtains the region segmentation figure of target image.

Step 2, carries out Edge contrast to former target image, obtains sharpening image, carries out Gauss and to be deteriorated the foreground extraction of model, obtain Gauss and be deteriorated a little in rgb space to sharpening image.

Step 3, is deteriorated a little filters the Gauss obtained in step 2, and acquisition Gauss is deteriorated key point p _{(x, y)}.

First our filter method is exactly first do filtration treatment to noise spot.Meanwhile, image is done to the process on de-black limit, then created a 3*3 filtrator and reduce candidate point.By the filtration of this step, will there is relative equilibrium in the distribution of candidate point.

Step 4, combines the result of step 1 and step 3, obtains the key point K being positioned at Ncut zone boundary _{(x, y)}.

The marginal information of image zones of different is found in the combination of the result that we are deteriorated by Gauss and Ncut, deletes the key point outside edge.After this step filtration treatment, we effectively can navigate to the edge of image.

Due to the performance issue of Ncut segmentation, the borderline region obtained is also unstable.To this, we take the mode of iteration, and the marginal information of Ncut segmentation under reservation different scale, the boundary information obtained like this can more stablize practicality.

Step 5, obtains the gonglion R of former target image _f.

This focus is positioned at foreground area under normal circumstances, centered by the geometric center of figure, according to nine grids model, a figure is equally divided into nine regions.Add up the number that the key point remained through a upper link drops on each region.By the comparison of the key point number to these nine regions, image focal point is positioned at the center comprising the maximum region of key point number in these nine regions.Initial focus is positioned at image geometry center.

Step 6, the gonglion weighted direction key point obtained in step 4 obtained in step 5 increases, and obtains final foreground area anchor point.

After step 5 obtains the focus of image, the focus direction making key point to image increases by we.Each image has four borders.Through analysis to great amount of images, find that the foreground area of most of image all can not be distributed on the border of image, if a panel region comprises many marginal informations simultaneously, so this panel region is that the possibility of foreground area is minimum.Therefore, add certain weight when we do increase process to each key point, the quantity on the border that weights and this region comprise is inversely proportional to.

Step 7, combines the foreground area anchor point of the region segmentation figure of the target image of step 1 gained and step 6 gained, obtains final foreground area.

Retain the cut zone that the foreground area anchor point number comprised is greater than threshold value, remaining region then thinks that background parts is cast out.

Accompanying drawing explanation

Fig. 1 the present invention is based on Gauss to be deteriorated the process flow diagram of display foreground extracting method of model;

Fig. 2 is specific embodiment of the invention procedure chart;

Fig. 3 is step 3 to the procedure chart of processing procedure of the key point obtained that is deteriorated to Gauss of step 6;

Fig. 4 is final sum additive method prospect comparison diagram.

Embodiment

Step 1, uses Ncut technology to carry out region segmentation to target image, obtains the region segmentation figure of target image.

Step 2, carries out Edge contrast to former target image, obtains sharpening image, carries out Gauss and to be deteriorated the foreground extraction of model, obtain Gauss and be deteriorated a little in rgb space to sharpening image.

We do Gaussian convolution to each pixel in image, and as the following formula shown in (1), G is Gaussian function, and I is the pixel in image, and we allow the pixel in image and Gaussian function do convolution.L is the matrix obtained after convolution.σ is the yardstick of Gaussian function, and we are 1.6 for the value of σ.

L(x,y,σ)＝G(x,y,σ)*I(x,y) (1)

In formula (2), we obtain Gauss and to be deteriorated difference original image being done to Gaussian convolution exactly under different scale of D, expression.

D(x,y,σ)＝L(x,y,kσ)-L(x,y,σ) (2)

We do Gauss's variation to each pixel in image under four kinds of different yardsticks, and result obtains this variation value of three floor heights.For each pixel, we are Gauss's variation value in its middle layer and eight points around it, and these eight point bilevel Gauss's variation values compare, if he is the extreme value of these 27 points, we are then deteriorated this point a little as Gauss.

Step 3, is deteriorated a little filters the Gauss obtained in step 2, and acquisition Gauss is deteriorated key point p _{(x, y)}.

Each Gauss is deteriorated a p _{(x, y)}value be this point of 1 expression be key point, value is this point of 0 expression is not key point,

$p_{(x,y)}=\left\{\begin{array}{cc}0,& S_{(x,y)}>1\\ 1,& S_{(x,y)}\≤1\end{array}\right.---\left(3\right)$

Wherein S _{(x, y)}represent at p _{(x, y)}the number that Gauss in the eight connectivity region at place is deteriorated a little.The be deteriorated point selected of Gauss is colouring information based on image, and change large pixel for color value and surrounding, the value that its Gauss is deteriorated also can be large.

Step 4, combines the result of step 1 and step 3, obtains the key point K being positioned at Ncut zone boundary _{(x, y)}.

For the key point K obtained in each step 3 _{(x, y)}, search in its eight connectivity region, if the institute in its eight connectivity region a little all with this point in same region, then remove this key point, otherwise, be then the key point of zone boundary this point location

K _{(x, y)}be that 1 this point retains, be 0 and this point is removed from key point.

Step 5, obtains the gonglion R of former target image _f.

Target image is divided into the region R of nine pieces of formed objects by the model of nine grids _i, the step 4 calculated in each block region obtains the quantity of key point, and the region that the Key number finding out step 4 acquisition is maximum, the central point in this region is decided to be the gonglion of target image

Initial focus is defined as the geometric center point R of target image.

Step 6, the gonglion weighted direction key point obtained in step 4 obtained in step 5 increases, and obtains final foreground area anchor point;

The weights W of key point p _pfor:

Key point is with W _pweight increase to gonglion direction.

Step 7, combines the foreground area anchor point of the region segmentation figure of the target image of step 1 gained and step 6 gained, obtains final foreground area.Experiment net result figure as shown in Figure 4.

Retain the cut zone that the foreground area anchor point number comprised is greater than threshold value, remaining region then thinks that background parts is cast out.Threshold value be chosen for 0.2, his implication is the ratio that foreground area anchor point quantity that this region comprises accounts for total foreground area anchor point quantity.

Claims (7)

1. be deteriorated based on Gauss the display foreground extracting method of model, it is characterized in that: when extracting display foreground, carry out following steps:

Step 1, uses Ncut technology to carry out region segmentation to target image, obtains the region segmentation figure of target image;

Step 2, carries out Edge contrast to former target image, obtains sharpening image.In rgb space, carry out Gauss to sharpening image to be deteriorated the foreground extraction of model, obtain Gauss and be deteriorated a little;

Step 3, is deteriorated a little filters the Gauss obtained in step 2, and acquisition Gauss is deteriorated key point p _{(x, y)};

Step 4, combines the result of step 1 and step 3, obtains the key point K being positioned at Ncut zone boundary _{(x, y)};

Step 5, obtains the gonglion R of former target image _f;

Step 6, the gonglion weighted direction key point obtained in step 4 obtained in step 5 increases, and obtains final foreground area anchor point;

Step 7, combines the foreground area anchor point of the region segmentation figure of the target image of step 1 gained and step 6 gained, obtains final foreground area.

2. the display foreground extracting method of the model that is deteriorated based on Gauss according to claim 1, is characterized in that: being deteriorated to the Gauss obtained in step 2 described in step 3 is a little carried out filter method and be, to be deteriorated a p for each Gauss _{(x, y)}value be this point of 1 expression be key point, value is this point of 0 expression is not key point,

$p_{(x,y)}=\left\{\begin{array}{cc}0,& S_{(x,y)}>1\\ 1,& S_{(x,y)}\≤1\end{array}\right.---\left(1\right)$

Wherein S _{(x, y)}represent at p _{(x, y)}the number that Gauss in the eight connectivity region at place is deteriorated a little.

3. the display foreground extracting method of the model that is deteriorated based on Gauss according to claim 1, is characterized in that: the border key point K described in step 4 _{(x, y)}acquisition methods be, for the key point obtained in each step 3, search in its eight connectivity region, if points all in its eight connectivity region all with this point in same region, then remove this point, otherwise, be then the key point of zone boundary this point location

K _{(x, y)}be that 1 this point retains, be 0 and this point is removed from key point.

4. the display foreground extracting method of the model that is deteriorated based on Gauss according to claim 1, is characterized in that: the acquisition target image gonglion R described in step 5 _fmethod be target image is divided into the region R of nine pieces of formed objects by the model of nine grids _i, the step 4 calculated in each block region obtains the quantity of key point, and the region that the Key number finding out step 4 acquisition is maximum, the central point in this region is decided to be the gonglion of target image

Initial focus is defined as the geometric center point R of target image.

5. the display foreground extracting method of the model that is deteriorated based on Gauss according to claim 1, is characterized in that: the method that the gonglion weighted direction that the key point described in step 6 obtains in step 5 increases is, the weights W of key point p _pfor:

Key point is with W _pweight increase to gonglion direction.

6. the display foreground extracting method of the model that is deteriorated based on Gauss according to claim 1, it is characterized in that: the region segmentation figure of the target image of step 1 gained described in step 7 and the foreground area anchor point associated methods of step 6 gained are, retain the cut zone that the foreground area anchor point number comprised is greater than threshold value, remaining region then thinks that background parts is cast out.

7. the display foreground extracting method of the model that is deteriorated based on Gauss according to claim 6, it is characterized in that: the region segmentation figure of the target image of step 1 gained described in step 7 and the foreground area anchor point of step 6 gained in conjunction with time, threshold value be chosen for 0.2, his implication is the ratio that foreground area anchor point quantity that this region comprises accounts for total foreground area anchor point quantity.