Font Size: a A A

Saliency Detection Algorithm Integrating Object Contour Information And Feature Contrast

Posted on:2018-01-06Degree:MasterType:Thesis
Country:ChinaCandidate:Z B ZhangFull Text:PDF
GTID:2348330533463261Subject:Electronic Science and Technology
Abstract/Summary:PDF Full Text Request
As performance of intelligent hardware improving and image process has been widely applied to various computer vision fields,real-time capability and precision of image algorithms are more focused.Saliency detection is similar to the selective mechanism of human vision,which provides computer with foreground and ignores background to obtain a high efficiency and accuracy rate.As an essential pre-processing step,this technique is widely applied to various computer vision fields,including image retrieval,image classification,and image segmentation,with the features above.First,the saliency ultimately aims to estimate the object position,and extract the usable information of the object.Hence,saliency computation is transformed into a probability inference problem in this object,which means that pixels belonging to object are more likely to be salient.Based on this theory,active contour analysis is combined with global color distribution under the framework of Bayesian model,to obtain a saliency model evaluating salient level by the probability.Compared with traditional convex hull,active contour can supply a more accurate contour for Bayesian model.Second,although the Bayesian model can predict the saliency distribution well,the probabilistic estimation contains many indetermination factors.Hence,for obtaining a more accurate saliency map,a novel model based on salient edges and remarkable discrimination of superpixel pairs is further introduced.A precise saliency map is then generated by ingrowth approach from these salient edges.Compared with probability estimation,edges are more applicability and credible as they are inherent attributes of images,and this model can suppress the feature-similarity problem between foreground and background.At last,since feature utilization of the saliency discrimination is not efficient enough,the results of the salient-edge extraction cannot satisfy us when processing some complex images.To improve the factor,therefore,a comprehensive model is proposed relying on sparse reconstruction with help of multi-layer contour zooming and multi-scale saliency revision.In this model,sparse reconstruction is used to detect object in single image for analyzing foreground and background features synthetically,which obtains a high-powered saliency model.
Keywords/Search Tags:global color distribution, active contour analysis, salient edges, sparse reconstruction, object contour zooming
PDF Full Text Request
Related items