Research On Saliency Detection Based On Multi-priors And Multi-scale

Saliency detection could quickly extract the most important parts of images and videos,which can effectively solve the problem of information redundancy and excessive calculation in images and videos.Therefore,it's applied in many scenes.On the basis of the study of various methods of saliency detection,this thsis focuses on the research of two hot models of boundary connectivity background priori saliency detection model(RBD model)and cellular automata saliency detection model(BSCA model).To solve their shortages,the detection algorithm is designed.The main research content is as follows.1)To solve the large error of calculating background probability caused by boundary connectivity in RBD model,a multi-priors saliency detection method combining background probability is designed.Based on the prior characteristics of background and foreground,boundary prior,weighted contrast prior,and target center prior are designed to detect image salicncy.On the basis of using the SLIC algorithm to segment the image into super-pixels,firstly from background,connected graph is constructed by affine propagation clustering and boundary priori is used to optimize background probability.And then from foreground,a coarse saliency map is obtained by weighted contrast prior and target center prior.Finally,the final saliency map is obtained by weighted minimization cost function and guided filter combining background probability and coarse saliency map.Experimental results on standard datasets ASD?ECSSD and SED2 demonstrate that the designed method has good effect compared with other mainstream methods.Compared with RBD model,the designed method has better P-R curve and MAE value decreased by 11.9%?6.1% and 6.8% respectively in three datasets.Adopting adaptive threshold segmentation,the designed method has a higher R value than RBD model in the case of comparable P and F values in ASD and ECSSD.Though P value of this thesis is slightly lower,both R and F values are much higher than RBD model in ECSSD,which can effectively highlight and smooth the saliency target.2)To solve the problems of significant background value and false detection at a single scale caused by BSCA model,a multi-scale background suppression method is designed.Firstly,the SLIC algorithm is used to segment the image into super-pixels at different scales based on multi-scale super-pixel segmentation.The global distinction map is constructed by clustering super-pixels of image boundaries and then a feature matrix is constructed by global distinction map,saliency map can be obtained by logistic regression.The logistic regression is used to obtain the saliency map.Then saliency map is updated by using the cellular automata synchronization update mechanism.Finally,the final weighted saliency map is obtained by color-weighted fusion.Experimental results on standard datasets ASD?ECSSD and SED2 demonstrate that the designed method has better effect compared with other mainstream methods.Meanwhile,compared with the BSCA model,the designed method has improved P-R curve in three datasets,and MAE value decreased by 12.9%,3.3% and 8.2%,respectively.Adopting adaptive threshold segmentation,the designed method has higher P,R,and F values than BSCA model in ASD and SED2.Though R value of this thesis is slightly lower,both P and F values are much higher than BSCA model in ECSSD,which can effectively increase the saliency of the target and suppress background.