Research On Image Saliency Detection Based On Attention Mechanism

With the help of the eyes and nervous system,humans can find valuable information in scenes in a short time.If computers have the ability to catch information quickly,they can serve human life effectively.Image saliency detection is to make computer have a human visual attention mechanism.The probability of pixel belonging to the foreground is calculated by using the features of image,and then the salient objects in the image are segmented.Based on human attention mechanism,this paper studies the bottom-up image saliency detection method.Static image saliency detection method is combined with motion information of video images,and then the saliency of multi-frame images in a video sequence is calculated.The main research content is as follows:(1)Aiming to solve the problem of blurred edges and internal unevenness in existing image saliency detection algorithms,this paper proposes a method based on undirected weight graph and multi-feature diffusion.First,an undirected graph is constructed with superpixels as nodes,and the connection mode of border superpixels is improved.On the basis of the improved graph,high-level feature is extracted using image color,texture features and prior knowledge of the image such as local contrast and center prior.Then the saliency map based on low-level features is acquired.Second,the maps based on foreground seeds and background seeds are calculated separately and fused using the high-level feature and compactness of salient objects.Finally,two-stage saliency maps are fused to acquire the final saliency map.Experiments show that the algorithm can clearly detect the edges of salient objects and can evenly highlight the entire salient region.(2)It's difficult to correctly detect salient objects in complex environment.To solve this problem,a saliency detection algorithm based on sparse reconstruction error and compactness of salient region is proposed.Firstly,the structure of the image is extracted to reduce background noise.Then it's segmented into several superpixels.On one hand,the background dictionary is formed by using boundary superpixels.Each superpixel is projected on the dictionary for sparse reconstruction to acquire a saliency map.On the other hand,the maps based on foreground seeds and background seeds are calculated separately and fused using compactness of salient objects.Finally,the saliency map obtained by the sparse reconstruction error and compactness are fused to acquire final result.Comparison experiments on data sets with complex images show that the algorithm has higher accuracy.(3)Aiming to detect the salient region of the sports scene,static image saliency detection method is extended to the video.A method is proposed to calculate the saliency of video images by using the static spatial features and motion features.Firstly,the image is segmented into several superpixels.On one hand,the background set is formed using static spatial features such as color and position of the superpixels,and then the background dictionary is built.Superpixels are sparsely reconstructed on the background dictionary,and then the saliency maps of video images are acquired.On the other hand,the motion features of the images are extracted.The background set is formed using motion features,and then the background dictionary is built.The saliency maps of video images are acquired using sparse reconstruction error.Then two stage maps are fused.Finally,the result is acquired by using the energy function to optimize the saliency map globally.Experiments show that the algorithm can effectively highlight the foreground and has a strong ability to suppress the background.