Image Saliency Target Detection Via Spatial Domain Feature Analysis

Visual saliency has always been a research hotspot in neurobiology,cognitive psychology and computer vision,it was originally used to predict the gaze position of the human eye in the scene,and researchers have found that people can quickly and accurately locate the most interesting part of the region from massive visual information,it can adaptively focus on regions of interest.Therefore,the image saliency detection algorithm is to simulates this mechanism and quickly obtains the salient target of an image,which improves the speed of image processing,so that it is popularized with other fields,such as target recognition and image classification,etc.Since the 1990 s,many image saliency detection algorithms have emerged,but with the improvement of image quality,some saliency algorithms can no longer meet the requirements of actual scenarios.In view of the results of saliency algorithm detection,there are problems such as missing edges,insufficient target brightness,inaccurate salient areas,and more background noise,we proposed two different saliency detection algorithms based on the spatial domain,the main research contents of this thesis are as follows:(1)We proposed a saliency detection algorithm via convex hull calculation and color features.Firstly,the channel values of images in different color spaces are different,we obtain a region contrast map by super-pixel segmentation in multi-color space.Then,the color boosted Harris method is used to form a convex hull to obtain a center prior map and a convex hull structure map.Next,a smooth channel difference map is obtained in the CIELAB space.The final saliency map is obtained by fusing four saliency map and optimizing them.(2)We proposed a saliency detection algorithm that combines random sampling and background priors.Firstly,in order to obtain the structural information of an image,the original image is divided into uniform super-pixel using SLIC algorithm in the CIELAB color space.Then,use square sub-windows with random sizes and positions are adopted on the three color channels of L,A,and B respectively,set a smaller number of sub-windows to updated iteratively and the Euclidean distance is used to obtain an integral random sampling saliency map.Next,use a selective mechanism to obtain reliable strong background seeds in the borders of an image and use them as the background regions to calculate the background prior map,and design a minimum energy function to optimize.Finally,the above two saliency maps are linearly fused.(3)Through subjective analysis of the two algorithms and 18 algorithms on the two image data sets of ASD and MSRA-B,and five kinds of objective evaluation indicators including precision-recall rate curve,automatic threshold segmentation,mean absolute error,structure measurement and computation speed are used to prove effectiveness of the two algorithms.Experimental results show that both algorithms can effectively suppress background interference,highlight salient regions with high brightness,and obtain full-resolution saliency maps with higher accuracy.