Research On Image Quality Assessment Algorithm Of High Dynamic Range Image Based On Deep Learning

The rapid development of mobile Internet and camera equipment is engendering an explosion of multimedia data.The way people directly express their opinions is transformed from information into images.On the one hand,high dynamic range images can reflect the real scene in detail,due to the limited dynamic range displayed by the display device,it is usually compressed by the image processing algorithm to the dynamic range of the output display device.It becomes one of the hotspots and focuses of high dynamic range image research.On the other hand,the subjective image quality assessment(IQA)has attracted a lot of attention in the past few years.Most existing traditional subjective IQA mainly focuses on predicting the mean opinion score(MOS),while ignoring the high diversity among people's perceptual quality.In order to solve the above problem,this paper proposes two methods of IQA via predicting score distribution.This paper proposes two models of high dynamic range image distribution of opinion score(DOS).The main work consists of following two aspects:(1)Subjective Image Quality Assessment based on Support Vector Regression:The proposed model incorporates feature extraction based on deep learning and regression prediction into a unified framework to simultaneously learn effective semantic features and accurate predictions.In particular,deep features in an image play a key role in representing global and local information;regression prediction which transforms from valid image information into opinion score distributions.We propose a novel IQA algorithm to predict DOS which can provide a more comprehensive information.The experimental results show that the proposed framework can effectively predict DOS.(2)An End-to-End Perceptual Image Quality Assessment Method via Score Distribution Prediction: In order to further promote the image feature transformation into the efficient and mutual learning of the opinion score distribution process,we propose to predict the distribution of opinion scores via an end-to-end convolutional neural network(CNN).The network is based on the depth residual network and introduces a new statistical region of interest pooling(SRP)layer to capture the global information and local details of the image,two fully connected layers are designed behind.To suit probability distribution learning,cross entropy loss(CE)is used as a loss function.Experimental results show that the proposed model can predict the distribution of opinion scores of images more efficiently and effectively.