Research On No Reference Image Quality Assessment For Realistic Distortion

In real life,people can evaluate the quality of an image without any reference information.The no-reference objective image quality assessment aims to automatically and accurately predict and perceive image quality scores without the reference image by designing a reasonable calculation model.However,there are still many difficulties in applying no-reference image quality assessment in real life.One of the most important points is that the research object of image quality assessment in the academic circles is mainly distorted images generated by artificial simulation.However,images captured using real-world mobile camera devices are usually afflicted by complex mixtures of multiple distortions,which are not necessarily well modeled by the synthetic distortions found in existing databases,thus limiting the wide application of no-reference image quality assessment in real life.In order to promote the image quality assessment methods in real life,the study of realistic distortion-free reference image quality is of great significance.This thesis focuses on the realistic distortion of image quality assessment issues.The main research contents are from three aspects: the traditional feature extraction method,deep learning method,and the industrial method of image quality assessment.A no-reference image quality assessment for realistic distortion based on structural statistical features is proposed.The algorithm describes the structure information of the image by extracting the characteristics of the total variation weighted local binary patterns,calculates the multi-scale information by combining multi-scale similarity,and uses the support vector regression to fit the mapping relationship between image features and human subjective evaluation scores,finally construct the no reference image quality assessment model.The experimental results show that the algorithm has high consistency for the objective prediction of synthetic hybrid distortion and human visual perception.For the photos with realistic distortions,this algorithm has better performance than other algorithms,but still has room for improvement.In the second part,a no-reference image quality assessment for realistic distortion based on deep learning is proposed.The algorithm has training and testing phases.The training phase first increases the training sample by dividing the image into image blocks and adding score labels to the image blocks through different methods in order to get enough training data.Then the image blocks and their tags are used to train the network to perform prediction of the partial images.In the test phase,the model inputs the complete test image,calculates its local image quality block scores,and then obtains the overall objective quality assessment score of the test image from the local image quality scores through different pooling strategies.The experimental results show that the algorithm has high consistency for the objective prediction of hybrid analog distortion and human visual perception.For realistic distortion image data,this algorithm is superior to other algorithms,but there is still room for improvement.In the third part concentrates on image quality assessment for realistic distortion based on Imatest.This method firstly uses the industrial image quality assessment Imatest software package to carry out the objective quality evaluation of the test equipment.By setting up a qualified test environment,selecting appropriate test indexes and corresponding test chart cards that conform to the standards,a mobile terminal camera test system is constructed,including resolution,geometric distortion,noise analysis,color accuracy,dynamic range,and image uniformity.From these aspects,an objective image quality assessment method is performed on the camera of the mobile terminal.Then subjective test to verify the subjective and objective consistency of the test standard.The experimental results show that the objective evaluation method of the mobile terminal is well consistent with the subjective evaluation method.