A Super Resolution Image Quality Assessment Method Based On CNN

With the advent of digital information age,various forms of data are displayed in front of people’s eyes and bring convenience to people’s life.Images have now become one of the main sources for people to obtain information,and they are becoming more and more important to people’s lives.At the same time,people want the image to be clearer,so that the message is richer.This requires image with high resolution,because of the limitation of physical equipment,cost,and technology makes the image in the process of imaging,acquisition or transmission will exist a variety of factors,which lead to image because of the different degree of degradation and distortion caused by the decline in the quality,the image is not clear,therefore,lost a lot of internal information.In addition,such low-resolution images often appear in our lives.Therefore,many researchers use software to process these low-resolution image information and convert it into clear and high resolution images with good visibility through signal processing or image processing.A large number of super-resolution reconstruction methods have been proposed,followed by the need to design a method to measure the performance of super-resolution reconstruction algorithm.At present,many methods for image quality assessment have been proposed,but there are few real algorithms for super-resolution image quality assessment.In this paper,the existing research methods in this direction are analyzed and their shortcomings are improved.The following is the whole work content of this paper:In this paper,a super-resolution image quality assessment method based on siamese neural network is proposed.Considering the insufficient number of database samples,the full-reference image quality assessment method is firstly used to input high resolution images and super-resolution reconstructed images into the network,and the convolutional feature extraction module is used to obtain the features of the input images to make full use of the information of the input data.In addition,wavelet transform is used to obtain the peak features of super-resolution reconstructed images and input them into the network model to improve the learning ability of the network in the case of insufficient data samples.In this paper,images are cut into 32×32 pieces to increase the number of samples in the database,and a method of label redistribution is proposed.Finally,a comparative experiment is conducted with other existing image quality assessment methods to verify the superiority of the proposed model.This paper also proposes an image quality assessment method based on bilevel recursive learning.Aiming at the problem that image information is not fully utilized due to the expansion of data set and image cutting.Firstly,the transfer learning method is used to make the network have the ability of feature extraction by fine-tuning the parameters of the network.The classical convolutional neural network is used as the backbone network,and the global multi-scale feature extraction module is combined to make the network model pay attention to the whole information and local distortion area of the image.At the same time,a two-level recursive learning method is designed to train on two data sets at the same time to enhance the generalization ability of the model.Finally,the effectiveness of the proposed method is verified by comparative experiments.