Salient Object Extraction For Three-dimensional Images And Segmentation Evaluation

Object segmentation algorithm and its evaluation is a key step of computer image analysis. It has an extremely important role and significance in image processing technology. With the rapid development of information technologies, more three-dimensional images appear in people’s life. Therefore, the research on salient object segmentation algorithm and its evaluation for the stereo images is listed as one of the key problems of image process. Saliency can simulate human visual attention mechanism, and can be expressed by a common feature in left-right 3D images, and this common feature, called disparity, which contains some inline features between the left and right image. Thus, when divide the left and right map and evaluate, stereo saliency will provide a good a priori knowledge.The three main contributions and innovations of this thesis are as follows:(1) Through the extraction and integration of disparity influence, central bias and spatially dissimilarity, we establish a new stereo saliency algorithm model, which can detect stereo image saliency regions more accurate. Disparity influence come from disparity map. We count gray probability density on a whole disparity map and then take internal product on gray probability level between patches. Meanwhile, we use a probability-weighted vector distance to compute disparity influence. Besides, we also integrates central bias and spatial dissimilarity to complement the part of the 3D saliency computation. According to the saliency likelihood of a patch should reduce with increasing distance from the center of the image, we extract central bias. In addition, we use a PCA-like method to reduce the image set. In the reduced space, we calculate a patch-based eigenvector difference as a measure of dissimilarity. These three factors are as basic judgment of saliency.(2) Bring the stereo saliency in object segmentation evaluation and develop experienced robot user for salient object interactive segmentation algorithm. We adopt an automatically balanced constraints based segmentation method to display our segmentation evaluation model. The segmentation method use graph-cut framework with the adaptive parameters of the two characteristics of the energy function. The adaptive parameters rely on the distance between image node to foreground and background in spite of manual selection. We establish an experienced robot user based on the concept of robot user and stereo saliency priori guidance. The experienced robot user used for segmentation evaluation which can completely independent simulate the process of human interactive segmentation. It only needs an original image and corresponding ground-truth and saliency map.(3) Generate two evaluation indexes: interaction quantity and error rate, then we build an evaluation model of interactive image segmentation. Based on above we do some experiments, the results have proved that the proposed method is feasible and effective, there is some value in use.