Blind Quality Assessment For Multiply Distorted Images

Images are introduced a variety of distortions in end-to-end compression and transmission,which results in its quality degradation.Therefore,it is necessary to perform Image Quality Assessment(IQA)for the received images when designers need to quantify the performance of multimedia compression and transmission systems.IQA has a long research history.At first,people mainly study single-distortion images containing only one type of distortion,but real-life images are often affected by multiple types of distortions.Hence,multiply distorted IQA has received attention in recent years.Compared to single distortion,multiply distorted IQA is more valuable and more challenging.In this paper,we focus on more challenging multiply distorted image quality assessment,and propose a MOS(Mean Opinion Score)training-based and a MOS training-free algorithm for the fact that the accuracy and generalization of multiply distorted IQA algorithm need to be improved.The training-based algorithm is used to solve the accuracy problem,and the training-free method is used to solve the generalization problem.In the MOS training-based algorithm,we propose a new idea for multiply distorted IQA.The new idea comes from the fact that we consider the damage caused by multiple distortions can be summarized as the detail loss and the detail redundancy,and the visual mask effect between them is harmful to IQA.According to this,we first decompose the distorted image into the detail loss map and the detail redundancy map,and design the feature representation for them respectively,then merge the two feature representations to form a new feature vector,and finally train the image quality prediction model based on SVR.The generalization of the MOS training-based algorithm is generally weak.Therefore,we propose a MOS training-free method with stronger generalization ability.This method is a difficultly complete-blind method in the field of quality assessment.It can work without MOS scores,which largely eliminates the time-consuming subjective quality evaluation work.The main working workflow of the algorithm is as follows: Firstly,we construct a multiply distorted visual codebook with known quality based on the perceptually relevant features.then calculates the weight of each visual word in the codebook,and finally weights the quality score of each visual word to infer the image quality score.The MOS training-based algorithm presented in this paper improves the average accuracy of the four public datasets by 3% compared to the current best algorithm GWH-GLBP.The proposed MOS training-free algorithm has an average generalization capability of 14% higher than GWH-GBP in the three multiply distorted databases.The experimental results show that the two algorithms have better consistency with the subjective evaluation of human eyes.