No-Reference Based Edge Enhancement And Feature Fusion Research Of Image Quality Evaluation Algorithm

No-reference image quality assessment aims to the research on automatic perception and quantification algorithm of image quality without manual intervention.As the main carrier of stream data,images would inevitably degrade due to compression and noise during transmission and storage.In addition,results of image fusion and generation in image processing requires image quality evaluation to measure whether it conforms to the perception of the human vision,therefore,accurate and efficient image quality evaluation is of great importance and value for both research and application.Although NR-IQA method does not require the label of reference image datasets,it also lacks reference indicators for manual evaluation of images.Therefore,we must use the information of the image itself for quality prediction.However,the existing NR-IQA method based on convolutional neural network has two main problems: 1)The convolution operation focuses on the local features of the image,but not the connection between image contexts.2)As the network deepens,it is apt to lose the detailed features of the image.According to the analysis of existing problems,in this paper the research focuses on the following two parts(1)A no-reference image quality assessment algorithm based on adaptive feature fusion is proposed.To address the problem above,the convolution operation is improved by expanding its receptive field,which could capture more associated context information and improve the discriminability of network on distortion features.To deal with second problem,the backbone network structure is optimized by fusing the features at different levels adaptively,which could obtain more detailed information.In addition,the image to be evaluated and its edge feature map are trained together to enhance the perception of image detail features.(2)A no-reference image evaluation algorithm based on boundary awareness and feature coupling is proposed.In order to further optimize the solution to the problems,a dual-backbone network model with CNN and Transformer feature coupling module is developed,so as to obtain the connection between local and global regions.In the coupling module,the local features and global representation are effectively combined in the manner of interaction,which compensates the loss of detail information to a certain extent.At the same time,we introduces a trainable Sobel convolution,which optimizes the extraction of image edge features in the previous work and improves the perception of image detail features.In summary,the main contributions of this paper is to improve the accuracy of quality prediction by capturing the regional information between images and enhancing the detailed features.A large number of qualitative and quantitative experiments verify the satisfactory performance of the model.The desired methods provide new ideas for optimizing the network structure design of NR-IQA,and image quality evaluation accordance for related researches such as image transformation and image generation.