Research On Super Resolution Reconstruction Method Based On Residual Network For Mine Image

With the rapid development of computer technology and hardware computing power,image super-resolution reconstruction technology,which aims to reconstruct low-resolution images into high-resolution images with rich texture and structure,has been widely used in various fields such as video surveillance,medical imaging,and satellite remote sensing.However,in the complex underground mining scenes,the images collected often suffer from severe problems such as dark blurring and unclear edges,which can have many adverse effects on the intelligent and safe development of the mining industry.Therefore,the study of super-resolution reconstruction of underground mining images is of great theoretical significance and practical value for the intelligent and safe development of the mining industry.At present,there are few studies on image super-resolution reconstruction methods especially for underground mining scenes.Existing image super-resolution reconstruction algorithms are mainly designed for normal environments and often fail to perform well when applied to mining images.Therefore,this thesis first constructed a coal mining underground image dataset CMUID for training and testing experiments of network models,which enhances the generalization ability of models in real mining scenarios.Then,this thesis focuses on the research object of mine images,and studies the super-resolution reconstruction method based on residual network in order to improve the ability of model feature extraction and model expression.The main research contents and results are as follows:(1)To address problems such as dark blurring and unclear edges in mining images,a super-resolution reconstruction method based on hierarchical features and attention mechanism for mine image is proposed.Firstly,a residual coordinate attention module is designed to incorporate coordinate attention mechanisms into residual blocks,allowing the network to obtain richer high-frequency detail information and better recover the edge detail information of mining images.Secondly,a hierarchical feature fusion mechanism is used to fuse feature information from different network layers,promoting the reconstruction of edge detail information and enhancing the reconstruction performance of the model for mining images.Experimental results show that the quality of reconstructed images using this algorithm is better than other comparative algorithms in terms of objective metrics and subjective perception.(2)A lightweight mining image super-resolution reconstruction method based on blueprint separable convolution was proposed to address the problem that current image super-resolution reconstruction methods increase algorithm complexity and memory consumption by increasing network depth and width,making it difficult to apply and deploy to actual edge mobile devices.Firstly,a lightweight residual attention module is designed,which uses blueprint separable convolution to replace the standard convolution in the residual coordinate attention module,and adds skip connections to improve the performance of residual block,allowing the model to maintain low parameter and computational complexity while also having good feature extraction ability.Secondly,an enhanced hierarchical feature fusion module is designed to perform local feature fusion followed by global fusion of feature information from different network layers,further promoting information flow in the network and enhancing the feature representation ability of the model.Finally,the pixel attention mechanism is used to improve the feature expression ability of the model and provide more important information features for the image reconstruction module.Experimental results show that this method has better reconstruction performance than other comparative algorithms with low computational and memory costs.