Salient Target Detection Based On Opponent Color And Pseudo-Wigner-Ville Distribution

Significant region detection is an important field of computer vision, has playedan important role in many important application fields, such as image segmentation,image retrieval, object recognition and adaptive image compression. With thedevelopment of Internet and multimedia technology, visual saliency also graduallybecomes a hot research topic. Visual saliency is a kind of perception, makes use ofhuman psychological adjustment mechanism, distributes computing resources priorityto those areas which we are more likely pay attention to, so as to improve theefficiency of image processing algorithms.Automatic target detection based on human visual selective attention is mainlycomposed of the generation of significant map and threshold segmentation. Thegeneration of significant map mainly has two kinds of types: one is bottom-up, basedon low-level vision information, using the underlying features of the image, thencombine into a remarkable map. The other is top-down, based on high-level visioninformation, make the salient map through empirical data of the image or naturalbackground.Most realistic computer graphics are color pictures, extract features using grayimages will lose part of the information, which affects the final recognition effect. Inthis paper, we make use of Opponent color space’s visual perception ability, proposesa new Four-Channel-Opponent-Color space, extracting color significance in differentscales. Unlike traditional color feature extraction methods, the model not only seeksfeature descriptors which represent target invariant, at the same time, based on thestructure properties of low-level visual perception, extracts the color, intensity andorientation information, the method is more matching the characteristics of humanvisual perception.While according to the Wigner-Ville Distribution(WVD) can showthe characteristics of image on the spectrum distribution, using anisotropic as theprinciple of image quality evaluation standard, studis the method for generatingsignificant map by adding one-dimensional window to Pseudo-Wigner-VilleDistribution(PWVD) and Renyi entropy, using super-pixel analysis to get the salienttarget. And on this basis, the principle of how Renyi entropy and pseudo wignerdistribution serve the anisotropy is given. For color images, especially the naturalscenes, the method can effectively detect the target with accurate results and goodrobustness. Based on the model of local characteristics expression, in this paper, acombination expression of local features and frequency domain features has beenproposed. First of all, using the feature extraction method of classical Itti model toextract orientation and intensity, while the color processing pathway is processed inthe Four-Channel-Opponent-Color space, which is add red/green channel to theclassic Opponent color space, then with the red/green, yellow/blue pathway togetherestablish a color salient map of local characteristics. Second, do one-dimensionalPWVD along the horizontal and vertical direction respectively, and then Renyientropy was calculated to get the sum of the two Renyi entropy, select the averageRenyi entropy as the final result. Then average filtering is done to get the frequencycharacteristics salient map. Finally, weight the characteristics of color, direction,intensity and frequency, using a simple joint operation and normalization to constitutea new significant map. On one hand, the method make full use of the orientation,intensity, color and frequency information of the image, enhance the reliability androbustness of the salient region. On the other hand, the proposedFour-Channel-Opponent-Color space is more matching the characteristics of humanvisual perception, using the high performance of changes in lighting conditions,increase the accuracy of image description. Finally, by using two kinds of evaluationcriterion, we do the experients on the MSRA1000database, the experimental resultsshow that the proposed method performs good, is better or equal to the five kinds ofclassic algorithms on all the criterions, can effectively meet the significant testingrequirements.