CN105023263A - Shield detection and parallax correction method based on region growing - Google Patents

Abstract

The invention provides a shield detection and parallax correction method based on region growing. Firstly an initial parallax graph is obtained via a tree structure dynamic programming algorithm, then a shield region and a region influenced by shielding are detected via a growing method of a doubt region at the edge of the initial parallax graph, and finally refilling is performed on the pixel point parallax values of the shield region and the region influenced by shielding by utilizing a color channel vectorization method so that a final parallax graph is obtained. Mismatching rate of shielding on the shield region and the region influenced by shielding can be obviously reduced.

Description

A kind of based on the occlusion detection of region growing and the method for parallax correction

Technical field

The invention belongs to technical field of image processing, be specifically related to a kind of based on the occlusion detection of region growing and the method for parallax correction.

Background technology

Eclipse phenomena is the study hotspot of binocular solid coupling always.Because the scenery degree of depth is different, distance video camera scenery far away is blocked by the scenery that distance video camera is nearer, thus can not be projected on video camera imaging face and form image.Namely under a certain viewpoint, visible scene areas may become occlusion area under another viewpoint, as shown in Figure 2.Eclipse phenomena causes larger error hiding to overall disparity map.The simple and effective detection occlusion area of existing certain methods energy, row are as parallax histogram peak rule of three (BMD), and the succession leash law (ORD) of matching double points, blocks leash law (OCC).These methods can only qualitative detection partial occlusion region, accurately cannot obtain occlusion area, simultaneously, these class methods are not all considered by blocking the region a little affected, the disparity map finally obtained has comparatively big error, coupling accuracy is low, so a kind of region-growing method that leaves a question open based on initial parallax image border of proposition of novelty determines the region being subject to eclipse phenomena impact, is called region of leaving a question open herein.

The regional processing that leaves a question open effectively can reduce the error hiding in this region.List of references (Q Yang, L Wang, R Yang, eta1.Stereo matching with color-weighted correlation, hierarchical belief propagation andocclusion handling [C], 2008) utilize the parallax probability of the geometrical constraint between Iamge Segmentation block and the pixel in block, estimate the optimum parallax value of incredible pixel.These class methods effectively estimate the parallax approximate value by blocking the pixel affected, but depend on algorithm segmentation result, so part better not realizing by blocking the disparity estimation a little affecting pixel, the parallax value of mistake being filled, causes increasing matching error.

Summary of the invention

The present invention proposes a kind of based on the occlusion detection of region growing and the method for parallax correction, obviously can reduce and block to occlusion area with by the error hiding rate of blocking range of influence.

In order to solve the problems of the technologies described above, the invention provides a kind of based on the occlusion detection of region growing and the method for parallax correction, comprising the following steps:

Step one: read binocular image I to be matched _land I _rif, left view I _lfor image to be matched, right view I _rfor matching image, obtain size and the Color Channel information of image;

Step 2: the dynamic programming algorithm using tree construction, obtains matching image I _rin each match point at image I to be matched _lthe point to be matched of middle correspondence, calculates shift differences and the parallax value of each match point and point to be matched, then parallax value is stored in initial parallax matrix, obtains initial parallax figure;

Step 3: detect the region of leaving a question open based on initial parallax figure according to the initial parallax matrix of step 2, described in region of leaving a question open be occlusion area and by the set of blocking the region affected;

Step 4: by Color Channel vectorization method, calculate in region of leaving a question open and leave a question open a little and each non-similarity left a question open a little in correcting window T, choose and a little the most similar non-pixel that leaves a question open that leaves a question open, using the non-pixel parallax value that leaves a question open as the parallax value left a question open a little, thus complete the parallax correction of parallax matrix;

Step 5: the parallax matrix corrected in step 4 is detected leave a question open region and parallax correction again, then step 3 is returned, repeat step 3 to step 5 calculating until in parallax matrix parallax point parallax value do not changing, terminate to correct, obtain the parallax matrix that correction of a final proof completes, parallax value in final parallax matrix is mapped to corresponding gray space [0,255], generates final parallax.

The present invention compared with prior art, its remarkable advantage is: (1) obtains initial parallax figure by tree construction dynamic programming algorithm, when extracting border to initial parallax figure, the error hiding reduced in background is also extracted into border, effectively reduces the interference of complex background; (2) detect occlusion area by the region-growing method that leaves a question open of initial parallax image border and blocked range of influence, not only consider occlusion area, consider the error hiding by blocking range of influence simultaneously, to occlusion area with blocked range of influence and all carry out parallax correction, effectively improve the coupling accuracy of algorithm; (3) Color Channel vectorization method is utilized to refill to occlusion area with by the pixel parallax value of blocking range of influence, the color data difference reduced between pixel is little, essence represents error hiding when different colours, than the single color similarity compared between the more effective judgement pixel of color difference; (4) the inventive method entirety does not need to depend on Iamge Segmentation, algorithm is comparatively simple, effective reduction algorithm complexity, counting yield far above above-mentioned based on the overall situation and the algorithm of Iamge Segmentation, effectively evade that to block the parallax caused fuzzy simultaneously, effective process is the information dropout caused because blocking the impact caused disparity map, improves the correct matching rate of disparity map.

Accompanying drawing explanation

Fig. 1 is the inventive method process flow diagram.

Fig. 2 be binocular solid coupling block schematic diagram.

Fig. 3 is the Teddy standard testing image that the Middlebury test platform of emulation experiment of the present invention employing provides, and Fig. 3 (a) is Teddy standard testing left view, and Fig. 3 (b) is Teddy standard testing right view.

Fig. 4 uses the initial parallax figure and final parallax that shown in the inventive method process Fig. 3, binocular image obtains, and wherein Fig. 4 (a) is initial parallax figure, Fig. 4 (b) is final parallax.

Embodiment

The present invention proposes a kind of based on the occlusion detection of region growing and the method for parallax correction, by Color Channel vectorization method, calculate and leave a question open a little and the non-similarity left a question open a little, choose and a little the most similar pixel that leaves a question open, using its parallax value as the parallax value left a question open a little, thus complete parallax correction.As shown in Figure 1, specifically comprise the following steps:

Step one: read binocular image I to be matched _land I _rif, left view I _lfor image to be matched, right view I _rfor matching image, obtain size and the Color Channel information of image;

Step 2: the dynamic programming algorithm using tree construction, obtains matching image I _rin each match point at image I to be matched _lthe point to be matched of middle correspondence, calculates shift differences and the parallax value of each match point and point to be matched, then parallax value is stored in initial parallax matrix, obtains initial parallax figure.The dynamic programming algorithm of tree construction is specifically see document (M.Bleyer and M.Gelautz, Simple but Effective Tree Structures for DynamicProgramming-Based Stereo Matching [C], 2008)

Step 3: detect the region of leaving a question open based on initial parallax figure according to the initial parallax matrix of step 2, described in region of leaving a question open be occlusion area and by the set of blocking the region affected.The present invention defines the outside of initial growth point, and namely " outward direction of growth " is by larger parallax value regional orientation comparatively disparity value region, for being blocked range of influence; The inside of initial growth point, namely " inward direction of growth " is by the comparatively larger parallax value region of disparity value regional orientation, is occlusion area.The process detected based on the region of leaving a question open of initial parallax figure is:

3.1. utilize canny operator to extract edge to initial parallax figure, obtain the boundary set based on initial parallax matrix, using the frontier point of boundary set as the initial growth point detecting region of leaving a question open.

Left pixel parallax value d in initial parallax matrix of 3.2 calculating growing points _lwith right pixel parallax value d in initial parallax matrix of growing point _rdifference, using the width that this difference grows left and to the right as growing point, so the difference of twice is width w (p) in region of leaving a question open corresponding to frontier point, the computing formula of w (p) is as follows:

w(p)＝2*(|d _l-d _r|)

The whole boundary set of 3.3 traversal, does same treatment to each frontier point, tries to achieve the region of leaving a question open of the whole boundary set of the correspondence after growth, the pixel gray-scale value in region of leaving a question open is labeled as 1, rest of pixels point gray-scale value is 0, is stored in and leaves a question open in matrix, complete the mark to region of leaving a question open.

Step 4: by Color Channel vectorization method, calculate in region of leaving a question open and leave a question open a little and each non-similarity left a question open a little in correcting window T, choose and a little the most similar non-pixel that leaves a question open that leaves a question open, using the non-pixel parallax value that leaves a question open as the parallax value left a question open a little, thus complete the parallax correction of parallax matrix.Detailed process is:

Leaving a question open of being labeled in matrix of determining to leave a question open is some pixel p to be corrected, according to color vector formulae discovery at image I to be matched _lpixel p to be corrected and correcting window T centered by pixel p to be corrected in non-other pixels leaving a question open concentrated between color cost value E _m, with color cost value E _mjudge to leave a question open in region and leave a question open a little and each non-similarity left a question open a little in correcting window T, color cost value E _mlarger, leave a question open a little and the non-similarity left a question open a little higher, color cost value E _maccount form be:

$E_M=\underset{p,q\∈T,q\∉\hat{o}}{\max }{\stackrel{\→}{v}}_p\·{\stackrel{\→}{v}}_q/\left(\right|{\stackrel{\→}{v}}_p|\·|{\stackrel{\→}{v}}_q\left|\right)$

Wherein, represent that pixel p to be corrected is at image I to be matched _lcolor Channel numerical value. the non-concentrated any pixel that leaves a question open in expression correcting window T is at image I to be matched _lcolor Channel numerical value.

The parallax value of the non-pixel that leaves a question open maximum for similarity is considered as reliable parallax, assignment gives pixel p to be corrected, terminate current operation and turn next pixel to be corrected, the pixel parallax value corrected is stored in new parallax matrix, realizing leaving a question open area filling so repeatedly.

Step 5: the parallax matrix corrected in step 4 is detected leave a question open region and parallax correction again, then step 3 is returned, repeat step 3 to step 5 calculating until in parallax matrix parallax point parallax value do not changing, terminate to correct, obtain the parallax matrix that correction of a final proof completes, parallax value in final parallax matrix is mapped to corresponding gray space [0,255], generates final parallax.

The beneficial effect of the inventive method can be further illustrated by following simulation result:

As shown in Figure 3, the Teddy standard testing image that the present invention chooses Middlebury test platform and provides is tested.Fig. 3 (a) is Teddy standard testing left view, and Fig. 3 (b) is Teddy standard testing right view.

Use the inventive method process binocular image as shown in Figure 4, Fig. 4 is obtained in step 2 of the present invention, Fig. 4 (a) is for calculating the teddy disparity map obtained, the teddy disparity map that Fig. 4 (b) completes for the inventive method by tree construction dynamic programming algorithm to the Teddy standard testing image that Middlebury test platform provides., as the border of the Little Bear in Fig. 4 (a) and the borderline region of toy house, there is obvious error hiding, and obviously improve in Fig. 4 (b), reduce the error hiding in this region in borderline region in Fig. 4 (a).Significantly can find out that the inventive method has good corrective action to the error hiding that eclipse phenomena causes.

Claims (4)

1., based on the occlusion detection of region growing and a method for parallax correction, it is characterized in that, comprise the following steps:

Step one: read binocular image I to be matched _land I _rif, left view I _lfor image to be matched, right view I _rfor matching image, obtain size and the Color Channel information of image;

Step 2: the dynamic programming algorithm using tree construction, obtains matching image I _rin each match point at image I to be matched _lthe point to be matched of middle correspondence, calculates shift differences and the parallax value of each match point and point to be matched, then parallax value is stored in initial parallax matrix, obtains initial parallax figure;

Step 3: detect the region of leaving a question open based on initial parallax figure according to the initial parallax matrix of step 2, described in region of leaving a question open be occlusion area and by the set of blocking the region affected;

Step 4: by Color Channel vectorization method, calculate in region of leaving a question open and leave a question open a little and each non-similarity left a question open a little in correcting window T, choose and a little the most similar non-pixel that leaves a question open that leaves a question open, using the non-pixel parallax value that leaves a question open as the parallax value left a question open a little, thus complete the parallax correction of parallax matrix;

Step 5: the parallax matrix corrected in step 4 is detected leave a question open region and parallax correction again, then step 3 is returned, repeat step 3 to step 5 calculating until in parallax matrix parallax point parallax value do not changing, terminate to correct, obtain the parallax matrix that correction of a final proof completes, parallax value in final parallax matrix is mapped to corresponding gray space [0,255], generates final parallax.

2. as claimed in claim 1 based on the occlusion detection of region growing and the method for parallax correction, it is characterized in that, step 3 is specially:

2.1. utilize canny operator to extract edge to initial parallax figure, obtain the boundary set based on initial parallax matrix, using the frontier point of boundary set as the initial growth point detecting region of leaving a question open;

Left pixel parallax value d in initial parallax matrix of 2.2 calculating growing points _lwith right pixel parallax value d in initial parallax matrix of growing point _rdifference, using the width that this difference grows left and to the right as growing point, so the difference of twice is width w (p) in region of leaving a question open corresponding to frontier point, the computing formula of w (p) is as follows:

w(p)＝2*(|d _l-d _r|)

The whole boundary set of 2.3 traversal, does same treatment to each frontier point, tries to achieve the region of leaving a question open of the whole boundary set of the correspondence after growth, the pixel gray-scale value in region of leaving a question open is labeled as 1, rest of pixels point gray-scale value is 0, is stored in and leaves a question open in matrix, complete the mark to region of leaving a question open.

3. as claimed in claim 1 based on the occlusion detection of region growing and the method for parallax correction, it is characterized in that, step 4 is specially: leaving a question open of being labeled in matrix of determining to leave a question open is some pixel p to be corrected, according to color vector formulae discovery at image I to be matched _lpixel p to be corrected and correcting window T centered by pixel p to be corrected in non-other pixels leaving a question open concentrated between color cost value E _m, by color cost value E _mthe parallax value of the maximum non-pixel that leaves a question open is considered as reliable parallax, gives pixel p to be corrected by its parallax value assignment, terminates current operation and turns next pixel to be corrected, the pixel parallax value corrected is stored in new parallax matrix.

4. as claimed in claim 3 based on the occlusion detection of region growing and the method for parallax correction, it is characterized in that, color cost value E _maccount form be:

$E_M=\underset{p,q\∈T,q\∉\hat{o}}{\max }{\stackrel{\→}{v}}_p\·{\stackrel{\→}{v}}_q/\left(\right|{\stackrel{\→}{v}}_p|\·|{\stackrel{\→}{v}}_q\left|\right)$

Wherein, represent that pixel p to be corrected is at image I to be matched _lcolor Channel numerical value. the non-concentrated any pixel that leaves a question open in expression correcting window T is at image I to be matched _lcolor Channel numerical value.