Research On Image Super-resolution Reconstruction Algorithm Based On Residual Learning

The environment,weather and other external factors will affect the quality of image acquisition and reduce the resolution of the image in the process of image transmission.Low-resolution images can affect people’s access to information,so it is critical to obtain information from high-resolution images.Usually software technology is used to improve the resolution of the image,and this image processing process is called image super-resolution reconstruction.Traditional methods cannot make full use of the information of low-resolution images.Introducing convolutional neural network into image reconstruction can extract richer deep feature information and obtain better reconstructed images.However,there are also many shortcomings.The network layer is too shallow and the ability to obtain information is insufficient,the receptive field is too small,and the network layer is too deep,which leads to difficulty in training,lack of low-frequency information,gradient disappearance or explosion,etc.These shortcomings will affect the quality of reconstructed images.For the disadvantages:This paper proposes two network models for image super-resolution reconstruction.The first is to use a parallel structure and introduce residual learning for super-resolution reconstruction;The second is to replace the conventional convolution with self-calibration convolution,and introduce coordinate attention to improve the self-calibration convolution module to reconstruct the image.According to the network structure,a super-resolution reconstruction algorithm combining parallel convolution and residual network is proposed.（1）A parallel structure is used to extract features.Two convolution modules with different sizes and structures are used to extract feature information,and then the feature information obtained by the two modules is fused,so that there will be more abundant feature information for reconstruction.Both modules use densely connected networks,and Concat the feature information of each layer in the module to prevent information loss and increase network circulation.（2）Using residual structure in the network structure.The idea of residual is introduced to improve the quality of the network.An adaptive residual network is added to the two modules of the parallel convolution respectively.An adaptive residual network is added to the two modules of the parallel convolution respectively.After updating the weights in the backpropagation through the adaptive residual network,the supplementary features can enhance the richness of the output features and overcome the disappearance of the gradient.The direct residual is introduced at the end of the entire network structure to improve the fidelity of the reconstructed high-resolution image,and the global residual can also supplement feature information.（3）Introducing a new loss function in training.The semantic feature of the reconstructed image is enhanced,and the perceptual loss is introduced.The perceptual loss is to constrain the original image and the reconstructed image at the feature level,and the reconstructed image can retain the higher-level semantic feature information in the original image.The perceptual loss can slightly improve the image quality,and the mean square loss can ensure good consistency between pixels,so the mean square loss and perceptual loss are used as the joint loss function.In order to suppress invalid information and improve feature utilization,an image super-resolution reconstruction algorithm based on improved self-calibration convolution is designed.（1）Extracting features using improved self-calibrating convolutions.Replacing conventional convolution with self-calibration convolution increases the receptive range of the self-calibrated convolution layer,prevents the introduction of useless information,and improves regional sensitivity.The self-calibration convolution splits the multi-channel feature map into two parts,retains the self-calibration operation part,and adds an attention mechanism to the direct convolution part,so that the region of interest can be located more accurately from the direction and position,and more the optimal position and spatial information make the reconstructed image quality better.（2）Using subpixel convolution reconstruction in the reconstruction part.Using sub-pixel convolution as the reconstruction module,the feature maps are recombined for reconstruction,which further avoids the addition of invalid information.A multi-scale structural similarity loss function is introduced,which better preserves high-frequency information,but loses color and brightness.L₁ loss can preserve color and brightness well,so L₁ loss and multi-scale structural similarity loss are used as joint loss.Compare the two models proposed in this paper with other models.In the objective evaluation,the PSNR and SSIM of the proposed algorithm in reconstruction are improved compared with the comparison algorithms;In subjective evaluation,there are better visual effects.