Research On Saliency Detection Based On Low Rank Sparse Decomposition And Markov Chain

Image saliency detection is one of the key research topics of machine vision.Because of its wide practical application value,it has become a research hotspot.However,how to extract useful information in images quickly and accurately is the key factor of image saliency detection.By saliency detection,the region of human interest is obtained,and then the target region and the background region of the image are separated,thereby realizing machine vision applications such as image segmentation,target recognition,and video tracking.This paper first introduces several classical saliency detection models,and compares the detection patterns and analyzes model characteristics of each model.Based on this,a saliency detection method based on low rank sparse decomposition and Markov chain is proposed.Firstly,the low-rank sparse decomposition algorithm is used to preliminarily image the image.Using the simple linear iterative clustering(SLIC)algorithm,the input image is segmented to obtain the super-pixel image of the image,and the color,texture and edge features of each super pixel are extracted and combined.High-level prior knowledge of images:position prior,semantic prior and color prior constitute the feature matrix of the image,using the robust principal component analysis algorithm(RPCA)to decompose the feature matrix into low rank matrix and sparse noise,low rank matrix representation the background area of the image,the sparse noise represents the salient region of the image,and the preliminary saliency map S0 is obtained;The binarization processing according to the obtained saliency map S0 can segment the super pixel node of the background region and the super pixel node of the saliency region;Then,based on the superpixel map,the Markov random walk graph model is established.The superpixel node of the background region is used as the background absorption node,and the saliency value of each superpixel is calculated by the absorption Markov chain,and the intermediate result saliency map S1 is detected.The superpixel node of the salient region is used as the foreground absorption node,and the absorption Markov chain is used to detect the result is a significance map S2;Finally,merge S1 and S2 get the final saliency map S.In order to prove the detection effect of this model,three image data sets of ECSSD,MSRA and PASCAL were selected for testing.The detection effect of the model was verified from two aspects:subjective and objective,and compared with the classical significant detection model.The experimental results show that the model is superior to other models in ROC curve and PR curve,which proves that the saliency detection model based on low rank sparse decomposition and Markov chain can detect the saliency region more accurately and has higher robustness.