A Study Of Image Super-resolution Reconstruction Method By Fusing Hierarchical Features And Dense Residual Connections

Image super-resolution reconstruction technology plays an important role in the field of computer vision,aiming to recover low-resolution images into high-resolution images with rich texture details by computer software methods,and it has been widely used in many fields.In recent years,with the continuous development of the information society,the computing power has been significantly improved,and deep learning technology has gradually received the attention of scholars,and many image super-resolution reconstruction algorithms based on deep learning have been derived,on the one hand,most of these algorithmic models only use a single network structure,which treats feature information indiscriminately by stacking convolutional layers,making it difficult to give full play to the optimal network performance of convolutional neural networks.On the other hand,many algorithms do not consider the influence of noise environment when reconstructing,and it cannot reconstruct low-resolution noisy images with super-resolution.The former is dedicated to improve the super-resolution reconstruction effect in ideal environment;the latter is dedicated to provide a solution for the super-resolution reconstruction task of low-resolution noisy images.The main work of the thesis is as follows:(1)To address the problem that the single-channel network structure cannot fully utilize the image feature information,resulting in low utilization of image feature information,this thesis proposes a fused hierarchical feature U-shaped network model.The model firstly designs a connection method that combines deep feature fusion with shallow residual connection,which fully fuses deep image features with shallow features and improves the utilization of feature information in the network,and the existence of residual connection reduces the training difficulty of the model and effectively prevents problems such as model degradation;secondly,a residual distillation attention module is innovated in the network to make the network more efficiently focus on the key features of the image,thus better recovering the detailed information of the image.The experimental results show that compared with eight comparison models,this model not only shows better objective evaluation indexes on four benchmark test sets,but also presents better reconstruction effects in terms of subjective visual effects.Specifically on the Set14 test set,the peak signal-to-noise ratio of the model’s 4-fold reconstruction results improved by 0.85 d B on average,and the structural similarity improved by 0.034 on average,which fully proved the effectiveness of the algorithm model.(2)In response to the fact that many super-resolution reconstruction algorithms do not consider the influence of the noise environment when performing super-resolution reconstruction of images,this thesis proposes a U-network model based on dense residual connections,aiming to complete the super-resolution reconstruction of lowresolution noisy images.The model firstly designs a denoising module based on dense residual connectivity to effectively denoise low-resolution noisy images by using the characteristics of residual learning;secondly,it extracts new feature maps from the denoised images by using the completely symmetric characteristics of the U-shaped network and feeds them into the residual distillation attention module,which can enable the model to recover the detail information of the denoised images.The experimental results show that compared with eight comparison models,the present model exhibits better objective evaluation metrics on all four noise-added test sets,specifically on the Noise-Urban100 test set,the 4-fold reconstruction results of this model have an average reduction of 0.219 in image perceptual similarity metrics and an average improvement of 0.117 in structural similarity.in addition,from the perspective of subjective visual effects,the The reconstructed images of this model basically eliminate the original noise points and present a better subjective visual effect,which fully illustrates that this model can achieve the dual purpose of denoising and reconstruction of low-resolution noisy images.