Research On Image Super Resolution Reconstruction Based On Attention Mechanism

Images are one of the main mediums for human beings to obtain information,the quality of images greatly affects the amount of information people obtain from them,the higher the quality of the image,the more information is provided.In real life,limited by imaging equipment,transmission means,environmental interference and other factors,high-definition images in some scenes are difficult to obtain,and it is necessary to rely on image super-resolution reconstruction,which is an image processing technology that converts low-definition pictures into high-definition pictures to obtain more detailed information.Image super-resolution reconstruction technology is a classic computer vision task,which is widely used in medical imaging,remote sensing imaging,video surveillance imaging and many other fields.Attention mechanism-based super-resolution image reconstruction model can use the attention mechanism to obtain effective texture information to make the reconstructed image more realistic.In this paper,through the analysis of current image super-resolution reconstruction models,we find that most models have problems such as face distortion and texture recovery unrealistic in the process of texture recovery.To improve the above problems,this paper designed an image super-resolution reconstruction model based on multichannel attention mechanism,which was studied from the aspects of model structure design and loss function selection.The main work of this paper included the following aspects:(1)A lightweight multi-channel attention mechanism was introduced,which used one-dimensional convolution for local cross-channel information interaction,and gave different weights to feature channels to focus on extracting more important image feature information,enhancing feature extraction and expression capabilities,and improving model effects.(2)A texture recovery module was constructed to enhance the structure of the model by introducing dense residual blocks to enhance the number of network layers while avoiding the disappearance of gradients.Batch normalization layers in dense residual blocks were removed and residual scaling was used to adjust the fused texture information adaptively to avoid artifacts in the reconstructed images,reduced the computational complexity of the model,and improved the quality of the reconstructed images.(3)In order to better transfer the texture of the reference images to the low-resolution images,a texture migration module was designed,which was divided into two parts,encoder and decoder,to locate and transfer textures,respectively,to enhance the accuracy of texture migration.(4)In the image super resolution reconstruction model,the loss functions can play a role in measuring the performance of the model.In order to preserve the spatial structure of the low-resolution image and make full use of the rich texture of the reference image to improve the visual quality of the generated image,three loss functions,namely,rebuilding loss,fighting loss and perceptual loss,were used to better play the performance of the network.The results of model rebuilding were compared and analyzed on four common test sets:CUFED5,Sun80,Urban100,Manga109.The model not only improved the objective evaluation index,but also effectively improved the visual effect of the generated images.