Research Of Saliency Detection And Fusion Based On Low Rank Theory And Space-constrained Coding

Saliency detection is a hot issue in computer vision.It aims at achieving saliency performance comparable to human visual system(HVS)and automatically detecting objects attracted by the HVS in an image.After nearly two decades of development,saliency detection has made great progress.Most saliency detection methods can handle well the images with relatively simple background and homogenous objects.However,detecting objects from complicated images still faces many challenges today.In this thesis,some new methods are studied in two directions,including saliency detection and saliency fusion.The main contributions are summarized as follows:(1)A double structured nuclear norm-based matrix decomposition model(DSN-MD)is proposed for saliency object detection.In the DSNMD model.a tree-structured nuclear norm(TSN)is firstly introduced to constraint both the bscckground and fore-ground regions.TSN provides stronger performance at capturing the underlying struc-tural information of the image regions including gobal structure,local structure,and internal structure of each node of the tree,and it deservedly inherits the advantages of both nuclear norm and sparsity-related norms.Moreover,high-level prior knowledge is integrated into the DSNMD model to enhance the saliency detection performance.Ex-periments demonstrate that the proposed method achieves superior performance with respect to previous unsupervised saliency detection models.and obtains comparable performance with supervised methods.(2)According to the basic pipeline of image classification,a space-constrained coding based top-down saliency detection method is proposed for generating class-specific saliency maps for natural images.This method builds a simple but efficient feature coding method LCCC by adding the spatial contextual information into the coding process.In LCCC.one feature is similar to another one only if they are with similar context,which is considerably discriminative for saliency detection.Moreover,a multi-scale based contextual pooling operation is presented to exploit the feature contextual information in multiple neighborhood scales.By incorporating LCCC and contextual pooling,the obtained feature representation has high discriminative power and hence benefits top-down saliency detection.Objectness cue are used to enhance the top-down saliency to obtain accurate saliency maps.Experiments show that the proposed approach provides more accurate saliency maps as compared to the previous work.(3)There are two conclusions by analyzing the results of the existing saliency detection methods.First,different saliency detection methods often behave differently over an individual images.Second,these saliency detection results often complement each other.Aiming to combine saliency maps from various methods,a robust principal component analysis(RPCA)based saliency fusion method is proposed.This method regards saliency fusion as a low rank matrix recovery problem,and uses RPCA model to recovery the low rank matrix.Experiments results prove that the proposed approach can effectively utilize the advantages of individual saliency detection methods and obtain a more accurate saliency detection result.Saliency fusion provides an interesting perspective for the research of saliency detection.(4)Following the RPCA model based saliency fusion,a double low rank based ma-trix recovery model(DLRMR)is proposed for saliency fusion.Similar to RPCA model,DLRMR also assumes that in the saliency feature space,the matrix corresponding to the background is low rank.The key difference between these two methods is that,RPCA assumes that the object matrix is sparse in the saliency feature space,while DLRMR maintains low rank for the object matrix.Experiments results prove that the DLRMR-based saliency fusion method performs better that the RPCA-based ap-proach.In addition,to reduce the computational complexity of the proposed fusion method,a saliency model selection strategy is proposed.