Image Quality Assessment Based On Local Texture Representation

With the development of image processing technology,the requirement for high-quality images are risen.Since images is polluted during the procedure of image acquisition,processing,compression,transmission and reconstruction,it is important to build up the image quality assessment to serve as quality monitor in this system.The evaluation standard of images is produced by human beings with subject image quality assessment methods.However,these methods are unbearable on waste of time and manpower,besides they are poor portability.Therefore,the researchers pay their attention on modeling the procedure of IQA automatically,and get predicted values that are in accordance with subjective values.One of objective image quality assessment method focuses on modeling the structure and characteristics of human vision system from bottom to up.With the incomplete understanding of HVS and the high computational complexity,this method cannot achieve good performance.In these days,the most state-of-the-art methods are based on extracting features from natural scenes statistics,especially for extracting local texture features.In this paper,two image quality assessment methods based on local texture representation are proposed.1)Image quality assessment based on multi-order local texture features extraction.Texture information is extracted in V2 of HVS,which is one of the low-level features.Moreover,these low-level features producing the sense of image quality are extracted from local regions.LBP and LDP operators can effectively extract textural features in local regions,which conforms to the characteristics of HVS.So they are applied to image quality assessment.LBP extract textural features via coding the direction of gradient information in neighborhood.In the meanwhile,LDP extract more detailed information with derivative pattern,such as turning point which is extracted by second order LDP.In view of the above facts,in order to increase the evaluation accuracy and the adaptability of various types of distortion,LBP,the second order and the third order LDP are employed to extract high order information.Following the feature fusion and quantification processing,the differences are mapped to objective scores.Experimental results on different databases and different types of distortion show that the objective scores obtained in this method are highly consistent with subjective perception.The proposed method not only has a high accuracy of image quality prediction but also is efficient in computation.2)Image quality assessment based on multiple local texture features extraction and synthesis.The perception of image quality mainly depends on the low-level features in HVS.And one of the typical features are texture feature.Since the earlier area in HVS extracts features in local regions,the local texture representation can be employed to describe quality-aware perception.In fact,these features are simple in structure and various in sorts.Single feature extracted in IQA can generally reflect very limited characteristics of HVS.As a result,methods based on single local texture feature are effective in some distortions,but cannot agree on other distortions that they are less sensitive to.Methods based on multiple local texture features extraction and synthesis can enhance the sensitivities to various types of distortions and improve the evaluation performance.In spatial domain,multi-order LDP operator can extract rich texture information with the less sensitivity to contrast,and Sobel operator completes the lack of texture feature from LDP.In transformation domain,local spatial-frequency feature extracted from log-Gabor to describe the texture information.In this paper,after feature selection and parameter optimization,an algorithm that employs the above-mentioned features are proposed.And then support vector regression is used to synthesis these features and to get image objective scores.In the experiments,the multi-feature algorithm improves performance.Compared with the similar algorithm,the prediction scores by this method has a good consistency with subjective values and is robust across different databases and various distortion types.