Image Weakly Supervision Saliency Detection Based On Image-Level

In computer vision tasks,the analysis and research of location points of interest to users is a long-standing and challenging problem.At the same time,with the emergence of various social media and the impact of new retail on traditional industries,people have huge processing requirements for image information,which makes video tracking,video surveillance,target detection and image segmentation tasks Using the saliency detection technology,it is seen that images play an important role in the information exchange between people in today's society.The image is another carrier of data.In a large amount of data,people are only interested in some of the information.Undoubtedly,after tens of thousands of years of evolution,the human eye can spontaneously and quickly find the part of the image that attracts attention.However,because the computer stores the picture differently from the human brain,it has more than enough to extract the saliency area of the picture insufficient.Before the deep learning came into view in 2012,the main method used by researchers to solve the problem of image saliency detection was to use contrast.It has been observed that there is a significant difference between the signific ant target in the image and the contrast between the region and the surrounding environment.In addition,since human visual nerve cells are more sensitive to the edge portion of the image,researchers have proposed a method of superpixel,which will Divided into one super pixel block.This method has a significant improvement over the previous method,but the effect is still unsatisfactory.The rise of deep learning in the past decade has been a major breakthrough in the field of artificial intelligence.Especially in the direction of image processing,the error rate of computer recognition of image categories is lower than that of human beings for the first time,which has triggered a wave of deep learning.However,the deep learning method requires a large amount of annotation data,and the label data required for the saliency detection of the image is a pixel-level annotation,which undoubtedly increases the cost of the research.Therefore,in view of the above problems,this paper proposes a new method for image saliency detection.Our method is divided into two stages: Stage 1: Firstly,using deep learning technology,using large-scale image-level label training data,extracting The foreground inference graph of the image,this method is called weak supe rvised training,which solves the problem of obtaining expensive samples of the training set while extracting the image features;Stage 2: Introducing the super pixel method in the traditional image processing to the original image The super pixel block processing is performed,and the foreground inference map obtained by extracting the image features in phase 1 is merged to achieve the purpose of further refining the significant target boundary.At the end of the paper,by comparing with other best algorithms in four benchmark datasets,the superiority of this algorithm in unsupervised algorithms is proved,and it has certain advantages compared with the full-supervised algorithm.