Research On Issues Of Image Quality Assessment Based On Visual Feature And Natural Scene Statistics

Image quality assessment plays a vital important role in image processing.It can be categorized into subjective and objective image quality assessment.Subjective image quality assessment is usually time consuming,expensive and cannot be embedded in a real-time image quality system,hence researchers have paid their attentions to design the objective image quality assessment metrics which are simple,reliable,efficient and effective.Based on the analysis of the limitation of the current quality evaluation methods and technical difficulties,we studied the feature selection and pooling strategy issues for full-reference image quality assessment and features extraction problem for no-reference image quality assessment(NR-IQA).Our work mainly includes the following parts:(1)A contrast and visual saliency similarity induced index for full reference image quality assessment is proposed.In the first place,it firstly computes the local contrast similarity map and global visual saliency similarity map.Then,the summation of deviation-based pooling strategy is adopted for the final quality score.The experimental results on three benchmark database(LIVE,TID2008,CSIQ)and remote sensing image database have shown that the proposed model is effective and efficient,which makes it suitable for real-time image quality assessment.(2)A general-purpose NR-IQA algorithm based on gradient magnitude similarity and natural scene statistics is proposed.The new method firstly extracts features from the point-wise statistics for single pixel values which are characterized by a generalized Gaussian distribution model and statistical features based on neighboring gradient magnitude similarity.Then a mapping is learned to predict quality scores using a support vector regression.The experimental results on the benchmark image databases demonstrate that the proposed algorithm correlates highly with human judgments of quality and have a good robustness and is computational efficient.(3)An image quality assessment model which combines perceptual features with natural statistics features is proposed.Four perceptual features-phase congruency entropy,mean phase congruency,mean gradient and entropy of the distorted images are selected in addition to the 36 natural statistics features of sharp patches.Support Vector Machine Regression(SVR)is adopted to build a relationship between image features and quality scores.Experimental results show that the proposed method accords closely with human subjective judgment.(4)A simple deep learning network,namely,PCANet induced general-purpose blind/no-reference image quality assessment(NR-IQA)is proposed.It firstly perform a contrast normalization,and then adopt PCANet-2 to automatically extract the features,eventually the Support Vector Machine Regression(SVR)is adopted to build relationship between image features and quality scores,yielding a measure of image quality.Experimental results demonstrate its effectiveness.