Important Visual Feature-based Image Quality Assessment And Image Compression

Among all information received by a human being, visual information is the main and the most important one. And the image is an important kind of visual information carri-er. From the acquisition, processing, compression, transmission, restoration to display of image or video, monitoring and controlling image quality are needed. Image quality as-sessment (IQA) can be categorized into two kinds:subjective ones and objective ones. The subjective one is the most reliable one (or the ground truth), but it is the most expensive, time-consuming, and can not be used in real-time system. Therefore, developing objective IQA metrics which can automatically, reliably and accurately predict the subjective image quality is meaningful. The traditional objective IQA metric such as PSNR is poor in predict-ing subjective image qualities. Although presently many alternative algorithms have been developed and also some achievements have been obtained, there is still no metric which has high prediction precision and reliability as well as linearly predicting subjective qual-ities at the same time. Considering the benefits and shortcomings of two main kinds of objective IQA metric (the human visual system based ones and the engineering based ones), in this paper, a best cut-in point (i.e., an important visual feature of image) is found. And a new IQA metric is built, which possesses the advantage of both kinds and avoids all their shortcomings, thus it can applied for practical usage.It has been found that image gradient is far more important than gray or color values in subjective image quality assessment. Several visual properties on image gradient have been found via this paper's research. First the Stevens'power law is also suitable for visu-al response to image gradient magnitude. Second, gradient direction really has important influence on subjective image quality. Other factors such as quality uniformity, the visual detection threshold of gradient, and visual frequency sensitivity also affect subjective image quality. Base on these findings, a full-reference IQA model by multiscale visual gradient similarity (VGS) is presented. The VGS model adopts a three-stage approach:In the first stage, two low-pass pyramids for the test image and the reference image are built and the im-age gradients for each scale are computed, then the global contrast registration for each scale is applied. In the second stage, pointwise comparison is given by multiplying the similarity of gradient direction with the similarity of gradient magnitude. In the third stage, intrascale pooling is applied, followed by interscale pooling; during this stage, the single image quality of each scale is multiplied by the contrast projection coefficient and the quality uniformity factor, as well as the contrast sensitivity function can be applied to the interscale pooling. The optimal values of two parameters of VGS are trained with existing IQA databases, and the good performance of VGS has been verified by cross validation. Experimental results on seven databases show that VGS is competitive with state-of-the-art metrics in terms of prediction precision, reliability, linear prediction, simplicity, and low computational cost. These achievements depend much on the intuition (easy to incorporate HVS properties) and simplicity of image gradients.Considering significant gradients mainly focusing on edges in image, to protect these primal gradients, an edge-based lifting scheme for image coding is proposed. The aim is to overcome blurring and ringing artifacts around edges with conventional discrete wavelet transform. The main idea is to divide an image into two parts:the image portions along edges and the rest of the image, then the one-dimensional edge-based wavelet transform (EWT) and the two-dimensional EWT are used for coding the two parts respectively. To improve coding performance, the preprocessing of the edges is an important part of the proposed approach. First we replace the narrow region between two approximately parallel and close edge segments by one middle line, second we prune the detected middle lines and edges by their smoothness to sharpness ratio. Experimental results show this method can achieve obviously higher visual quality when compared with JPEG2000, especially at low bit rates.