No-Reference JPEG image quality assessment based on Visual sensitivity

in this paper, a novel human visual sensitivity measurement approach is presented to assessment the visual quality of JPEG-coded images without reference image. The key features of human visual sensitivity (HVS) such as edge amplitude and length, background activity and luminance are extracted from sample images as input vectors. SVR-NN was used to search and approximate the functional relationship between HVS and mean opinion score (MOS). Then, the measuring of visual quality of JPEG-coded images was realized. Experimental results prove that it is easy to initialize the network structure and set parameters of SVR-NN. And the better generalization performance owned by SVR-NN can add the new features of the sample automatically. Compared with other image quality metrics, the experimental results of the proposed metric exhibit much higher correlation with perception character of HVS. And the role of HVS feature in image quality index is fully reflected. Index Terms Human visual sensitivity; support vector regression; neural network; image quality; No-reference assessment I INTRUDUCTION Image quality evaluation plays an important role in processing image. With the extensive application of image, developing image quality metric without reference image has received widespread attention especially when it is difficult to obtain reference image. Thanks to image serving people, image quality assessment is more and more dependent on the characteristics of human visual system (HVS). Considerable volume of research has demonstrated that image quality evaluation methods considering human visual characteristics is better than others not considering these characteristics [1]. Therefore, it is imperative to develop the no-reference image quality metric based on human visual factors. In the last few decades, extensive valuable research has been carried out in developing this topic. Gastaldo et al. proposed a circular back propagation (CBP)-based image quality evaluation method [1], Venkatesh Babu et al. proposed a no-reference image quality index using growing and pruning radial basis function (GAP-RBF) [2] and Suresh et al. proposed a no-reference metric based on extreme learning machine classifier [3]. JPEG is one of the most popular and widely used image formats in internet and digital cameras. In this paper, for JPEG images, the extracted visual sensitivity approach is used to assessment the visual quality of images without any reference. The key human visual sensitivity factors were used as input vectors of network. Image quality estimation includes computation of functional relationship between HVS features and subjective test scores. Here, the functional relationship is approximated using support vector regression neural network. The experimental results show that the proposed no-reference image quality metric has a good consistency with mean opinion score (MOS), really embodying the role of HVS features in image quality measurement. Zhang You-Sai and Chen Zhong-Jun done the research work together in Jiang Su University of science and technology, and were grateful for LIVE database supported by Prof. H.R. Sheikh, etc. Email: yszhang100@163.com; chen.magic@163.com 46 No-Reference JPEG image quality assessment based on Visual sensitivity Copyright © 2011 MECS I.J. Modern Education and Computer Science, 2011, 1, 45-51 HVS II -BSED FEATURE EXTRACTION A key distortion of JPEG images is horizontal and vertical blocking artifact generated by DCT-based transform coding for per 8×8 image block. In order to measure this kind of distortion accurately, several important human visual sensitivity factors such as edge amplitude and length, background activity and luminance [4] are taken into consideration. Given a gray scale image of size N M × . The intensity of the image at any pixel location ) , ( j i is given by ) , ( j i I . The algorithm is explained for extracting human visual features along horizontal direction. (1) Edge amplitude and length: Edge amplitude quantifies the strength of edge along the borders of 8×8 blocks; Edge length quantifies the length of continuous block edges. They are obtained by horizontal orthogonal sobel filter operator. ⎩ ⎨ ⎧ < = others P I P I E e , 0 | * | |, * | τ , (1) P is the sobel horizontal filter, E is edge information. Threshold e τ should be below 40, if choosing a lower threshold might result in missing the real blocky edges resulting from compression. (2) Background activity: Background activity is denoted by the amount of high frequency texture content around the block edges. It is extracted by following method: ⎩ ⎨ ⎧ < = others F I M a ah h , 0 | * | , 1 τ (2) Here, ah F is high-pass filtering, h M is background activity. The value of threshold a τ in our experiments is 0.3, the effect of blockiness is masked by the activity if a τ is more than 0.3. Background luminance: Background luminance (3) measures the amount of brightness around the block edges. It is obtained by following way: ⎪⎩ ⎪ ⎨ ⎧ ≤ ≤ = others j i I fl I j i W p l , 1 128 ) , ( 0 , 128 * ) , (