Research On Image Quality Assessment Algorithm Based On Deep Learning

In the current information age,digital image,as an important carrier of information communication,has always played a very important role during the communication.However,due to the existence of some objective reasons,the image in the image processing system will lose part of the information and cause the image quality to decrease,which affects the subsequent image processing work and human reception of image information.The image quality assessment is to study how to evaluate the image quality changes,is a very meaningful research topic,and is also an important part of the image processing system.The image quality assessment algorithm can be divided into 2D image quality assessment algorithm(2DIQA)and stereoscopic image quality assessment algorithm(SIQA)according to the different types of images.As the name implies,these two types of image assessment algorithms act on flat and stereoscopic images,respectively.According to different evaluators,the image quality assessment algorithm can be divided into subjective image quality assessment algorithm and objective image quality assessment algorithm.The subjective image quality assessment algorithm is for people to evaluate the image quality according to their subjective feelings,while the objective image quality assessment algorithms is that the machine obtains the quality of the image by establishing an assessment model.The paper introduces the image quality assessment algorithms at first,and analyzes the challenges faced by 2DIQA and SIQA.In the field of 2DIQA,with the development of deep learning in recent years,many image quality assessment models based on deep learning have appeared,but the existing models are prone to overfitting problems in a small data volume environment.In the field of SIQA,the disparity and other stereo vision characteristics will greatly affect the feature extraction during the evaluation process.In view of the problems in the above two fields,this paper proposes an improved DIQa M?NR image quality assessment model and a stereo image quality assessment model based on shift-convolution.1)An improved DIQa M?NR image quality assessment model: this model uses transfer learning to improve the DIQa M?NR model.The improved assessment model uses Res Net50 network structure to replace the feature extraction layer in the original model,and transfer out the parameters of Res Net50 network on Image Net data set.Finally,this paper uses the global average pooling layer(GAP)to replace the fullyconnected layer(FC-512)in the original model.The improved assessment model has lower parameters and a deeper structure than the original model.Experiments show that the improved assessment model can simulate the human visual system and evaluate quality of image accurately even in a small data volume environment.2)Stereoscopic image quality assessment model based on shift-convolution: This model first uses several convolutional layers to extract low-level features of the stereoscopic image,then uses shift-convolution to layer to build a matching candidate set,and calculate the preliminary disparity information of the stereoscopic image.After that,the model uses multiple convolutional layers as optimization functions to optimize the candidate set and extract more accurate disparity information.In addition,existing research shows that the features of the left and right images also have a significant impact on the quality evaluation of stereoscopic images.Therefore,the model uses two sub-networks to extract features related to the perceived quality of the stereo image from the left and right images.In the feature mapping section,this paper introduces the GAP layer to reduce the dimensionality of features.Finally,the model linearly splices the dimensionality-reduced features,and uses a fully connected layer to learn the mapping relationship between these features and the quality of the stereoscopic image,and outputs the quality of the image.Experiments show that the accuracy of this model is higher than that of most assessment models.