No-reference Image Quality Assessment Based On Feature Extraction In Transform Domain

With the rapid development of multimedia technology, digital image which plays a very important role in communications has been widely used in all aspects of people’s life and work. However, processes from image acquiring, various types of processing, image compression to image transmission, may produce distortion in visual quality of digital image. So how to effectively assess image quality is becoming more and more important. Digital image quality assessment can be usually divided into subjective assessment and objective assessment, and objective assessment includes three methods, which are full reference, reduced-reference, and no reference image quality assessment. The research of this paper is about no-reference image quality assessment and the main work of this paper can be summarized as follows:1. Describes the research background and significance, analyses the image quality assessment methods both home and abroad, discusses the advantages and disadvantages of subjective image quality measurements, full-reference image quality measurements, reduced-reference image quality measurements and no-reference image quality measurements.2. Analyses the advantages of neural network, studies the basic concepts and theoretical knowledge of generalized regression neural network (GRNN) and fuzzy theory, and then proposes the idea that we can establish the fuzzy generalized regression neural network (Fuzzy-GRNN) by applying fuzzy theory with the general regression neural network.3. The paper puts forward blind image quality assessments based on Contourlet transform. Firstly, by studying the Contourlet transform we find that it can be used to capture the geometric characteristics of images. Based on the theory that distortions will destroy natural image statistics features, we extract image energy features by Contourlet transform and then combining with neural networks to assess image quality. Feeding the neural network with texture features and difference average subjective score (DMOS), we can obtain no reference image quality assessment models. Meanwhile, for the unsuitable distortion, we make amendment. Experimental results show that Contourlet transform based methods can effectively evaluate a variety of distortion types of image by giving a very consistent result with subjective evaluation. And fuzzy GRNN algorithm is better than GRNN algorithm in the overall performance.4. By studying the basic principle of nonsubsampled Contourlet transform (NSCT), we learn that it can better capture image geometric features than Contourlet transform. Based on this point, the paper puts forward NSCT based no-reference image quality assessment which can better reflect distortion degrees. The method firstly feeds energy characteristics and DMOS into neural network to establish a model, and then uses the learned model to predict image quality scores. Experiments show that this model correlates more highly with human visual perception on many distortion types than Contourlet transform based ones.