Visual Saliency Detection Based On Multi-Scale And Multi-Channel Mean

With the development of Internet and multimedia technologies, a large number of images have been provided by web server,which brings a great challenge for image processing and analysis. Humans routinely and effortlessly judge the importance of image regions, and focus attention on important parts, exhaustive study of saliency detection that allows the computer has the ability to capture salient has a far- reaching significance in improving the performance of image understanding and image analysis systems, as well as in enhancing the level of application of image processing technologies.This paper aims at improving the definition and accuracy of visual saliency detection. Based on attention mechanism of human visual system and component, frequency, channel color, size, etc. image information, we explore the saliency detection technology in several areas and propose a series of visual saliency detection method. The main work are listed as follows:(1) An approval for image saliency detection based on multi-component mean is proposed. Gaussian filter and color space conversion are used to increase image sharpness and filter background noise. Principal component transform is used to eliminate the interference of random factors and reduce the amount of image features need to be processed. The salient is calculated under three principal components. Finally,linear mean fusion is employed to obtain the final saliency map.The experimental results show that the proposed method can significantly improve the accuracy of salient detection.(2) An approach for image saliency detection based on multi-scale and multi-channel mean is proposed.2-D wavelet transform is used to decompose and reconstruct image, where can effectively filter the background information of image and highlight salient regions without image segmentation.The bicubic interpolation algorithm is used to narrow the filtered image in multi-scale.We take the distances between the narrowed images and the means of their channels as saliency values,where can avoid tending to produce higher saliency values near edges instead of uniformly highlighting salient objects. In order to filter the background noises of saliency maps, we only reserve part values which are not less than the mean saliency of a given image. Bicubic interpolation algorithm is used again to amplify the images in multi-scale, and then the saliency map is calculated by adding the amplified images. Finally, linear normalization is employed to obtain the final saliency map.Experimental results demonstrate the proposed method performs excellently in terms of definition and accuracy.(3) Extension experiments are proposed based on (1) and (2) respectively. Under the detection framework of (1) and (2) respectively,we expand color space model, color channels and principal components of an image, and detect visual saliency in different color space models,different principal components and different channels. Experimental results show that extended experiments can be more clear, more complete extraction of the salient areas, but also significantly improve the image visual saliency detection accuracy and relevance.