Saliency Region Detection Based On Diffusion Model With Statistical Optimizing

As the improvement of science technology and the development of Inter-net, all kinds of information has increased dramatically. As a major part in human life, image plays a more and more important role in information trans-fer. Therefore, computer vision has become an integral part of computer sci-ence. As for the problem that how to extract the valid and important informa-tion from images quickly and accurately, image saliency region detection has become an important part in computer vision as well as a hot topic. Saliency region detection simulates human visual attention mechanism, aiming at detect the most attractive region in every image, which named as saliency region. Im-age saliency region detection is always worked as the pre-processing for other areas of computer vision and now is widely used in image retrieval, image cut-ting, target detection, image compressing, etc. All in all, further researching in image saliency region detection is very significant for image processing, im-age understanding, and even the whole artificial intelligence area as well as the Internet development.Through the efforts of many researchers, saliency detection has developed into a relevant theory, and many state-of-the-art algorithms have been proposed.However, there are also some problems: 1. Saliency region detection usually based on some prior knowledges which are obtained by people defining. This may cause negative effects to detection result because of the lacking of flexi-bility. 2. Current algorithms usually pursue high precision but ignore recall,which may lead the salient object cannot be emphasize perfectly.In response to these problems, this paper do some research from following aspects:1. Objectness is introduced to obtain the prior background seeds in dif-fusion model to enhance flexibility as well as cover more scenes, so that to improve the accuracy of saliency detection results.2. Cellular automata is introduced to construct an iterative updating mech-anism, which is used to refine the foreground and background through the in-teraction among nodes.3. Bayes statistical model is introduced for optimizing in pixel level to deal with the problem that inaccuracy of superpixels cutting and ignorance of some details. In this part, salient value of each pixel can be calculate based on the previous result with statistical models, which can consider each detail and avoid image cutting to improve both precision and recall.Finally, in order to examine the validation of this proposed algorithm, other 9 classic or state-of-the-art methods are used to compare. The comparison will be done in 6 different datasets with 4 measurement indexes, so that to make the experiment more convenient.