Image Quality Assessment Based On Visual Perception And Neural Network Algorithms

Images are playing an indispensable role in our daily life and many research fields.With the advances in computer and network technology,digital image processing technology has developed rapidly.Image quality assessment(IQA)is desired to evaluate the perceptual quality of an image in a manner consistent with subjective rating,which is of great research significance.Perceptual properties of human visual system(HVS)is necessary to be considered during the design of IQA algorithms for their important role in subjective assessment.Owing to the great improvement in computer memory capacity and computing speed,deep learning methods increasingly perform well in image processing and computer vision researches.Based on the knowledge of HVS properties and deep learning theory,this paper researches into two-dimensional image quality assessment and stereoscopic image quality assessment,and the main research contents and achievements are listed as following:1.A full-reference image quality assessment method that employing joint special and frequency domain feature extraction is proposed based on the properties of HVS.Firstly,the saliency maps of reference images are obtained and used to enhance the reference and images.Then,the log Gabor filters and local tetra patterns are employed to capture spatial frequency and local texture features from the enhanced images,respectively.And then similarities between the features are synthesized and mapped into an objective quality score by regression scheme.Experimental results show that this method achieves higher performance comparing to other representative IQA methods,and is robust across different databases.2.A no-reference quality assessment method for stereoscopic images using convolutional neural network for adaptive feature extraction is proposed.A deep learning framework is introduced in this method,and a convolutional neural network is trained and finetuned to adaptively extract features from left and right view of a stereoscopic image pair,instead of design hand-crafted features used in conventional works.By fusing the features of two views and combing the complementary statistical features extracted from disparity map,the final high dimensional features are obtained.Then the features are mapped to subjective quality scores using a regression model,and the quality assessment is implemented.Experimental results demonstrate the superior performance of the proposed method over other existing methods,in terms of its accuracy in predicting stereoscopic image quality as well as its robustness across various databases and distortion types.