The Research Of Image Segmentation Algorithm Based On Visual Salient Region

Image segmentation is a basic and key step in advanced image processing operations such as object analysis,recognition and target tracking,etc.The quality of segmentation results directly affects the subsequent processing operations.Existing image segmentation algorithms can be divided into interactive segmentation and automatic segmentation according to human participation or not.The former has been widely used because of its good segmentation performance.However,when faced with a large amount of image data,the corresponding needs to consume more manpower and time-consuming,unable to meet the actual demand.In recent years,the research based on visual salient region detection algorithm has achieved good results,many researchers combine it with existing image segmentation algorithm,providing more possibilities for automatic segmentation algorithm.However,the visual salient region detection algorithm is still unable to effectively detect salient regions in images containing uneven gray distribution,complex background and foreground similar to background features(such as gray level or color),which affects the image segmentation accuracy.Therefore,this paper focuses on the research of image segmentation algorithm based on visual salient region.The main work and innovations of this paper are as follows:(1)A region-based automatic initializing active contour model(initial contour based on visual salient region-region based hybrid active contour model,ISR-RBHM)is proposed.Firstly,the improved visual salient region detection algorithm is used to preprocess the segmented image and access to object candidate regions,the initial contour curve is set automatically;secondly,the obtained object prior information is combined with the bitmap with the maximum contrast in the image to be segmented,the adaptive symbol function is designed,and the optimized Lo G(laplacian of gaussian,Lo G)energy term is weighted,which is fused into the RSF(region-scalable fitting,RSF)model in a linear way to enhance the adaptive ability of the model;finally,a new local grayscale measure is designed,which is combined with the local kernel function to improve the local energy term,improving the sensitivity of the model at the weak edge of the object,and accurately locate the object boundary.The experimental results show that the ISR-RBHM model can effectively identify visual salient region,automatically set the initial contour and effectively retain the details of the object edge.The visual and quantitative experimental results prove that the ISR-RBHM model is superior to some mainstream active contour models at present.(2)Aiming at the problem that the existing visual salient region detection algorithm cannot better identify the foreground object region in the image containing complex background,so as to reduce the segmentation accuracy.Firstly,a two-layer sparse graph is constructed,and then a saliency map based on compactness is calculated to locate the object region more accurately;secondly,the superpixel of the image boundary was selected as the background seed,and the saliency map based on background seed was proposed.The SRCB(salient region detection algorithm based on compactness and background seed,SRCB)algorithm was combined with the compact saliency map in a linear way,which effectively detected the saliency object region and suppressed the background noise to a certain extent;finally,the fused saliency map are fused into the graph cut framework to eliminate the redundant background information in saliency map and obtain more accurate binary segmentation results.As an automatic segmentation algorithm,SRCB-GC(salient region detection algorithm based on compactness and background seed-Grab Cut,SRCB-GC)algorithm can effectively segment the foreground object region and remove the background noise without manual intervention.Experimental results show that the algorithm has obvious advantages when processing the images with complex background.(3)A cosegmentation algorithm for RGBD image is proposed.Firstly,a two-layer sparse graph is constructed,and then depth information is introduced to optimize the edge weight,the saliency maps based on foreground compactness and background seeds are obtained by depth information optimization,the saliency map corresponding to a single image in image group are obtained by linear combination;secondly,the initial foreground seeds in a single saliency map were selected,then the initial seeds were screened by combining the depth,color information,which can obtain the optimized foreground seeds,it used for constructing the foreground dictionary for sparse reconstruction;furthermore,the global reconstruction saliency map based on sparsity is calculated,at the same time,the relationship between the two images in the image group is fully utilized to construct the sparsity based local reconstruction saliency map,and the sparsity based global and local reconstruction saliency map are combined in a linear way,the fused sparse reconstructed saliency map was combined with the saliency map corresponding to a single image in a multiplicative way,and the fused saliency map was optimized by using energy functional to obtain the optimized saliency map;finally,the graph cut is used to segment the saliency map.It is not difficult to see from the experimental results that the algorithm achieves more satisfactory segmentation results and can satisfy the human visual sensory system better.