Research On Stereo Matching Algorithms By Visual Saliency

Stereo vision is an important researching direction in the computer vision.Stereo vision has been widely used in surveillance, human computer interaction, 3D measurement, robot navigation and 3D stereo film etc. In which stereo matching is an important and key researching field, since the accuracy and speed of stereo matching directly influence the performance and applications of stereo vision. The common of stereo matching algorithms choose characteristic from the image itself unilaterally. There is often exist some short comings, such as low speed, low efficiency, regardless the importance of pixels. Our researching is with reference to the mechanism of human active vision, in which the model of visual saliency is introduced to stereo matching. We presented and improved a series of stereo matching algorithms based on visual saliency. By doing so, not only the image characteristics inherent in the objective are considered, but also the active vision of human eye could be considered. The proposed algorithms can improve the matching accuracy and save the time of matching. The main contents are as follows:(1) A new visual saliency detection method for color image was presented. It uses the color information of image, the Hyper Fast Fourier Transformation(HFFT)and the accumulating of multi-sale saliency map to locate the focus area. A novel image fusion method is presented. Depends on pyramid decomposition and new saliency detection method, it can get a better performance in the experiments.(2) A stereo matching algorithm based on the saliency points was proposed, in which only the points with high saliency are used for matching. The matching criteria is maximum probability after moving the high saliency points to overlapping.The proposed method has higher efficient and can highlight the salience area than the method base on SIFT in compared experiments.(3) Local stereo matching based on SAD(Sum of Absolute Difference) is the highest effectiveness algorithm. However, the traditional algorithm based on SAD sensitive to the illumination variation. In order to solve this problem, an improved local stereo matching algorithm based on saliency map was presented. The methodutilizes the pixel features such as Sobel, phase and saliency of image. By means of image saliency, we can not only locate the main target, but also overcome the effect of illumination variation. By combined with Sobel and phase features, we can located the edge of main target accurately. Experimental results show that the proposed method could achieve a better disparity map compared to the traditional methods.(4) Adaptive weighted stereo matching algorithm has the highest accuracy rate among local stereo matching algorithm. However, it also exist some shortcomings.Such as incomplete descriptions of pixels weight, similarity measure sensitive to light changing and object occlusion. In order to solve these problems, we propose a new adaptive weighted algorithm for stereo matching. In which, RGB and gray value of pixel describe the color and brightness of the pixel, coordinates describe the spatial position of pixel, and saliency describes the significant of pixel. We can calculate the weight of pixel according to those pixel features. A new ratio rule is proposed to measure the similarity between pixels. Finally, we can compose the fusion image from left and right disparity map based on cost aggregation to solve the occlusion. The proposed method performs better than compared methods in the experiments.