Research On Image Super-resolution Reconstruction Method Based On Attention Mechanism And Wide Activation

Image super-resolution reconstruction is a very important research direction in the underlying computer vision tasks,whose main task is to reconstruct the corresponding high-resolution image based on one or more low-resolution images of the input.Super-resolution reconstruction techniques play an important role in improving the accuracy of recognition and classification in advanced computer vision tasks such as target recognition,image retrieval and image classification.Compared with traditional reconstruction methods,the effect of image super-resolution reconstruction based on convolutional neural networks is significantly improved,but there are still problems such as ignoring the spatial feature information of low-resolution images,incomplete extraction of high-frequency information,and large number of network model parameters.To address the above problems,this thesis analyzes and optimizes the network structure and residual module,etc.The main contributions of the dissertation are as follows:(1)An Image Super-resolution Reconstruction Based on Attention and Wide-activated Dense Residual Network(WDRN)is proposed for the problem of insufficient spatial feature information extraction and low utilization of low-resolution.Firstly,considering that the single-size convolution kernel cannot fully extract the shallow feature information of the low-resolution image,and the lost spatial features directly affect the extraction of deep features,the reconstruction model uses four parallel convolutional layers of different scales to fully extract the global and local information of the low-resolution image in the shallow feature extraction part,and then inputs the extracted information into the spatial feature conversion layer as the a priori information to guide the high-frequency.Secondly,in the deep feature mapping module,a dense residual group is constructed by incorporating lightweight attention into the wide-activated residual block as the basic unit for extracting deep features,which enriches the diversity of high-frequency features while adaptively adjusting the weights of key information in high-frequency features.Finally,a network structure with global residuals and local dense connections is constructed to enhance the forward propagation of network features and Improving feature utilization(2)A Dual Path Image Super-resolution Reconstruction Based on Wideactivation and Feature Distillation Networks(DFDN)is proposed to address the problems of a single feature extraction method of network structure and possible information redundancy in the feature fusion stage.First,the model builds a two-way parallel network structure and introduces global residual connections by using the residual network as the backbone network to improve the utilization of shallow features and accelerate the convergence speed of the network;second,it uses a residual feature distillation block with fast training speed and small number of parameters and incorporates the attention mechanism and wide activation idea to further enhance the ability of extracting high-frequency features;finally,to avoid the disadvantage that the elements may cause information redundancy due to undifferentiated summing,a gated fusion mechanism is introduced to learn different weights for the feature information extracted from the two branches separately to obtain the weighted fused features and thus improve the reconstruction performance.To verify the effectiveness of the algorithm in this thesis,the reconstruction results are experimentally compared with several mainstream deep learning-based algorithms under Set5,Set14,BSD100 and Urban100 test sets ×2,×3,×4magnifications.The results show that the number of parameters of the two network models proposed in this thesis is maintained at a low level,and the reconstructed images have clearer edge contours and richer texture detail information in subjective visual perception.