Region Of Interest Coding Based On Saliency Detection

With the development of the Internet and big data,massive amounts of image and video information is transmitted through the Internet every day.This puts tremendous pressure on data compression,transmission and storage.Therefore,image coding becomes one of the hot topics in information technology today.Traditional coding methods are devoted to removing the redundancy of information by removing the correlation of the original image,while ignoring the visual redundancy caused by the visual characteristics of human eyes.As a result,in this paper,we proposed a region of interest coding method which combines visual saliency with traditional coding.It can guarantee the subjective quality of images even when the bit rate is low.This paper mainly focuses on the study of saliency detection and region of interest coding.The main research are summarized as follows:(1)Depth information provides the distance from the scene to the camera,which can help to distinguish the foreground from the background area.In this paper,we use the depth information to extract the foreground of the color image.Then,the color image and the extracted foreground part are taken as two layers respectively.We can detect the salient region on the first layer by taking full advantage of the color information,while on the second layer by the color and depth information.The saliency maps of the two layers have their own advantages and disadvantages.In order to complement each other's advantages,central prior knowledge and background prior knowledge are utilized to obtain the better saliency map;(2)During the research,we found that most images on the Internet are typically stored in the compressed domain.There is great practical value in calculating the saliency information of these images.However,existing saliency detection algorithms have to decompress these JPEG images from the compressed domain into the spatial domain,so that it has a very extensive amount of computations.Therefore,in this paper we present an effective saliency detection algorithm,which is mainly based on DCT and secondary quantization.Firstly,the DC coefficient and the first five AC coefficients are used to get the color saliency map.Then,through secondary quantization of a JPEG image,we can obtain the difference of the original image and the quantified image,from which we can get the texture saliency map.Finally,the final saliency map is generated based on these two maps and two priorities;(3)With the extraction of salient region,we consider a region of interest coding scheme,so that saliency map can be extracted without complicated image decoding.The proposed method derives the relationship between the quality and saliency of images through the rate-distortion function.Subsequently,we encode the image by transforming the QP(quantitative parameter)value of the macro blocks to achieve region of interest coding,then we can get the saliency map through the QP value at the decoder.In order to evaluate the performance,we proposed a new criterion of image subjective quality called saliency region-weighted PSNR(SPSNR).Experimental results demonstrate that the proposed method can improve the whole SPSNR performance,especially for a high PSNR gain for visual important regions.Most important of all,the saliency map can be recovered directly and accurately.