Research And Application Of Saliency Computing Model Based On Stereo Vision

The 21st century is the age of information explosion, the information processing technology becomes the focus of attention. Time complexity and space complexity will be inestimable, if the same method is used to process all the information. The study shows that in the vast amounts of information, the information what is needed is just a small part of all the information. With the popularity of stereo imaging equipment, this thesis introduces disparity to study the stereo saliency detection, proposing data filtering of related task and extracting the interesting information to process. The saliency research has wide applications in the field of computer vision and machine learning, for example, it is applied in image search, target recognition and video compressing and so on. The study of this thesis mainly includes two aspects:In the first place, the thesis builds a saliency model based on the analysis of disparity to detect salient regions of the stereo images. Firstly, disparity information and region information is got by means of the preprocessing of the stereo images. After that, color complexity and disparity complexity is computed according to color feature and disparity information. Then the two kinds of saliency based on region can be aggregated. In the end, as we know the comfort zone which has zero disparity is the most salient region, and the background changes little in the disparity contrast in every row. It should be taken into consideration. Then the local contrast of disparity should be computed to increase the saliency of the objects and restrain the saliency of the background. The thesis achieves higher precision and recall ratio on the stereo images data set.Secondly, this thesis proposed an improved algorithm of the image seam carving which used the saliency algorithm. Consequently, images of different lengths-width ratios can be got to match different display devices. Firstly, a new energy image should be computed to keep cutting lines away from the salient region to avoid losing the important contents. After that, a new tailoring criteria is proposed and an energy function is computed. By this method, not only cutting lines in the image edges is scattered to refrain from the question of edge distortion which is caused by cutting lines clustering, but also the question is resolved that non-adjacent region created new edges when removed carving seam. The proposed seam carving algorithm not only stores the principal information, but also makes the image edges continuous and has a good image visual effect.