Image Saliency Detection And Application Based On Multiple Priors And Graph Model

Image saliency detection uses the computer to simulate visual attention mechanism of the human eyes,and automatically extracts the regions of interest from the images,which is the basic research of image processing.The existing saliency detection algorithm detects a large amount of background noise when the foreground and background colors are similar.When the foreground is located at the image boundary,the detected salient targets is missing.This thesis proposes an image saliency detection algorithm based on multiple priors and graph model.The main work and innovations of this thesis are as follows:Image saliency detection based on multiple priors and graph model.Calculate multiple prior information based on superpixels: objectness prior information,texture prior information,color prior information,boundary prior information,and salient prior information,which can be accurately detected when the foreground and background color are similar.A fully connected graph model is constructed,and the saliency map is calculated based on different boundary prior information,and the detected salient targets is well kept at the boundary portion.At the same time,optimization is performed in a variety of ways,including based on color and position information optimization,optimization based on fully convolutional network(FCN),and optimization using boundary refinement network(BRN)to consistently highlight salient targets.This thesis implements the application of image saliency detection in the field of cosaliency detection and co-segmentation.When the background of the image is complex and there are multiple foreground targets,the common foreground targets in the same group is difficult to detect completely.For this problem,an image co-saliency detection algorithm based on collaborative item optimization is proposed.By using collaborative item to optimize saliency maps,common foreground targets can be fully detected.On this basis,this thesis proposes an image collaborative segmentation algorithm based on global item and local item.The global-based data item,local-based data items and smoothing item are used for cosegmentation,and the common foreground targets in the same group are completely and accurately segmented.The image saliency detection algorithm based on multiple priors and graph model proposed in this thesis is subjective and objective contrast experiment with the state-of-theart algorithms on MSRA-10 K,ECSSD,SED2,PASCAL-S and DUT-OMRON datasets.The experimental results show that the proposed algorithm outperforms the existing algorithms in PR curve,F-measure and S-measure.The image co-saliency detection algorithm based on collaborative item optimization and the image co-segmentation algorithm based on global item and local item are compared with mainstream algorithms on i Coseg dataset.The experimental results demonstrate the superiority of the proposed algorithm.