Study Of Image Quality Assessment Based On Visual Features

Image is subject to distortions during acquisition, processing, coding, stor-age and transmission, which is caused by acquisition system, storage medium,processing algorithm, transmission equipment and so on. To maintain, controlthe quality of image, it is important for image system to be able to identify andquantify image quality degradation. Therefore, assessing image quality automat-ically and accurately is a meaningful research topic. Subjective image qualityassessment (IQA) by the direct evaluation from people can give the most accu-rate quality assessment results. However, such methods are not only expensiveand cumbersome, but also unable to be incorporated into the automatic imagesystem. Therefore, it is desirable to design objective quality metrics which canautomatically assess the quality of images in agreement with subjective scores.This thesis investigates visual feature extraction and application in IQA. We fo-cus on the single scale, multi-scale and statistical feature extraction, the designof feature similarity metric and pooling strategy for full and reduced referenceIQA.Firstly, we propose an edge feature based full reference IQA method. Sincehuman eyes are highly adapted to extract structural information, we extract thevisual feature from edge pixels, which contain the most important structuralinformation of image. It is observed that many types of distortions can not onlychange the gradient magnitude but also the orientation. Both of them can leadto the change of structural information in edge, and make people perceive thedegradation of quality. Thus, we propose a Magnitude-weighted Edge OrientationHistogram (MEOH) as the quality feature to describe the local distribution ofedge gradients as well as directions simultaneously. Then we define the distance metrics to calculate the bin-by-bin and global similarities between two MEOHdescriptors. The average pooling strategy is used to pool the overall quality scorefrom quality map. Our algorithm can be applied to a variety of distortion typesand degree of distortion. Experimental results verify the efectiveness of qualityfeature as well as the overall algorithm.Secondly, we investigate the multi-scale feature extraction, and propose amulti-scale wavelet feature based full reference IQA method. We use the wavelettransform to emulate the multichannel behavior of the human visual system(HVS). We introduce the two-dimensional wavelet leader pyramids extractedfrom wavelet coefcients to robustly extract the multi-scale information of edges.Based on the wavelet leader pyramids, we further propose a visual informationfidelity metric to evaluate the single-scale quality similarity, then we adopt ascale-variant weight approach to combine various single-scale quality similaritiesinto the final quality score. The weight for each scale is determined by the corre-lation coefcient between single-scale quality score and the subjective evaluationscore. In this way, the middle frequency sub-bands have higher weight than thelow-and high-frequency. This is more consistent with the property of HVS. Ex-perimental results demonstrate that our algorithm systematically outperformsstate-of-the-art IQA methods both in accuracy, robustness, and computationalefciency.Finally, we investigate the statistical feature extraction for reduced reference(RR) IQA. We propose a new RR IQA algorithm based on the image statisticsand visual saliency. We use the natural scene statistics (NSS) model to fit thegradient distribution of image in pixel domain locally. The model parametersare estimated as the quality feature, since they are highly correlated with HVS.We further define the similarity between the original and the distorted qualityfeatures as the local quality score. Finally, we propose a saliency based poolingmethod. We establish a natural image database, and collect the eye movementswhen navigating the images. We extract the quality feature of salient regionsfor the whole database, and estimate the feature probability density functionthrough a statistical learning method. Image regions with higher feature prob-ability means more saliency. In the pooling stage, the pooling weight for each block is the probability of the corresponding quality feature extracted from thereference image. Therefore, there is no need for the reference image to provideany additional information. Our quality feature has low data rate and the sim-ilarity metric has low computation complexity. Experimental results show thatthe proposed RR IQA method is more consistent with the subjective measurethan other well known RR IQA metric and the widely used full reference IQAmetric Peak Signal to Noise Ratio (PSNR).