No-reference Image Quality Assessment Based On Mass Image Data

Image data usually suffer distortion inevitably in the process of acquisition,storage, transmission and processing and this makes it hard to understand andexploit the image information. Obviously, the work of image quality assessment isan essential work. Since the human beings are the ultimate observers of the images,so the human perceptual quality is regarded as the criterion of image qualityassessment methods. The process of perceptual quality estimation which is namedsubjective image quality assessment is very expensive and time-consuming. As aresult, it cannot directly used in real-time system. So a lot of researchers devotethemselves to the work of objective quality assessment. The objective assessmentmethods can be classified into full reference (FR) ones, reduced reference (RR) ones,and no reference (NR) ones according to whether the reference image is available.By the way, the no-reference image quality assessment approaches are the mostwidely used ones.This paper proposes a general-purpose no-reference image quality assessmentmethod: we first retrieve numerous highly correlated images in a large imagedatabase as reference. Then we exploit the success of full-reference methods toestimate the quality of the degraded image. By a large number of experiments, theproposed method can overcome the drawbacks of existing general-purposeno-reference methods and delivers a highly consistency with human subjectiveevaluation.In general, the main contribution of this paper can be summarized as thefollowing three aspects:(1) An image alignment method based on genetic algorithm is proposed here.Through the extraction and matching of SIFT feature points, we can achieve a plentyof SIFT matching pairs. Then we will implement genetic algorithm to find thecorrect four matching pairs and calculate the correct transformation matrix. Finally,we can achieve the correct image alignment.(2) We have given the experimental results of the performance of themainstream block matching metrics under various noise situations. Due to thedegradation image in this application, so here we need to analysis the stability ofblock matching metrics and choose the best one to finish the patch matching work.(3) A novel mass image data based no-reference distortion metric is proposedhere. The most existing no-reference methods have several limitations as bellow: a)These methods always need a large human scored image database for training. b)Training data dependence. c) They cannot provide local quality map. As inspired by the success of the processing of massive image data, we put forward using fullreference method to solve the no-reference problem. So our method canfundamentally avoid the limitations of existing regression model based no-referencemethods.