Research On Lightweight Image Super-Resolution Reconstruction Model Based On Deep Learning

Image super-resolution reconstruction is a classic low-level computer vision task with extremely wide applications,such as remote sensing,security,monitoring,and medicine.Image super-resolution reconstruction refers to the reconstruction of highresolution images with good visual effects and more detailed information from corresponding low resolution images that have lost high-frequency information.In recent years,with the rapid development of deep learning,it has been widely used in image super-resolution reconstruction.Although deep learning based image superresolution reconstruction models have achieved good performance,the large number of parameters in the model makes it difficult to apply to resource limited devices such as mobile phones and embedded devices.In response to the above issues,this article aims to design a lightweight image super-resolution reconstruction model based on deep learning:1.This paper proposes a lightweight image super-resolution reconstruction model IRN(Image super-resolution reconstruction based on Conv Ne Xt Residual Network)based on Conv Ne Xt residual structure.Specifically,based on the traditional Conv Ne Xt residual structure,research and improve the residual structure suitable for image superresolution reconstruction tasks.(1)Construct residual blocks by cascading one standard convolution and multiple improved Conv Ne Xt residual structures;(2)Introduce the enhanced spatial attention module into the residual block to improve the representation ability of the model;(3)Stack multiple residual blocks to form the backbone network of the IRN model and fuse the sub-pixel convolution module to realize the mapping from shallow level features to deep level features.Finally,experimental analysis on five publicly available benchmark datasets showed that the proposed model IRN achieved competitive results in evaluating indicator values.2.This paper proposes a lightweight image super-resolution reconstruction model BCRN(Blueprint Separable Convolution and Conv Ne Xt Residual Network)based on blueprint separable convolution and Conv Ne Xt residual structure to address the issues of large parameter quantities and insufficient utilization of channel term information in the IRN model.Specifically,(1)constructing residual blocks by cascading 1blueprint separable convolution and 1 improved Conv Ne Xt residual structure to learn lost high-frequency information;(2)Introduce a hybrid attention module that enhances spatial attention and contrast perception channel attention into each residual block,enabling the model to better learn valuable features;(3)Stacking multiple residual blocks is used to form the backbone network.Dense connections are used between residual blocks to improve the utilization of features,and sub-pixel convolution module is introduced to realize image reconstruction.Finally,experiments have shown that BCRN has better results in both qualitative and quantitative evaluation indicators compared to most other advanced lightweight models.Compared with the IRN model,the BCRN model reduces the number of parameters and improves performance.