Research On Key Technologies Of Image Super-resolution Reconstruction In Visual Monitoring Of High-speed Rail Infrastructure

With the rapid development of China’s high-speed railways and the continuous improvement of vehicle operating speed,it puts forward higher requirements for the safety of railway infrastructure.The computer vision-based monitoring system plays an extremely important role in ensuring vehicle safety and high-speed operation..However,the visual imaging system is affected by complex factors such as hardware resolution and harsh natural environment,resulting in insufficient resolution of the collected monitoring images and degraded quality,which seriously affects the monitoring effect and brings hidden dangers to driving safety.The use of image super-resolution reconstruction technology to restore low-resolution monitoring images to high-resolution images has important application value for improving the accuracy and timeliness of high-speed rail infrastructure visual monitoring and ensuring the safety of high-speed rail operations.In this paper,the urgent need for non-contact visual monitoring of high-speed rail infrastructure is based on deep learning theory.For the high-speed rail infrastructure visual monitoring image details,especially the problem of poor reconstruction of geometric features or edge features,the deep convolution network is The main research object,starting from the loss function and network structure,to improve the image quality of infrastructure visual monitoring,research on super-resolution reconstruction method for high-speed rail infrastructure visual monitoring image.The main research contents of this paper are as follows:(1)An image super-resolution algorithm using generative adversarial network with edge guidance is proposed.Inspired by the game theory between the generator and the discriminator of the generative adversarial network,considering the edge feature prior to monitoring the image,the edge detector is used to add the edge loss to the generator.After fully training the network,the generator can more fully learn the edge features of the image,and the generated geometric texture of the super-resolution image is clear and prominent.The effectiveness of the super-resolution reconstruction method combined with edge features is verified by experiments.(2)An image super-resolution algorithm using dense skip connection networks with fusing low-level features is proposed.The algorithm proposes a new super-resolution network framework,and introduces a dense skip connected image generator to propagate low-level network feature information to the entire generator network,so that the network can effectively learn texture and edge low-level features.The super-resolution reconstruction of low-quality,fuzzy and unrecognizable visual monitoring images can be carried out,and the effectiveness of the dense hopping connection network with the underlying features is verified through experiments.(3)An integrated network image super-resolution reconstruction algorithm combining denoising functions is studied.The algorithm draws on the idea of integrated network,adopts the integration technology to perform feature-level weighted fusion of the sparse coding network and the dense hopping connection network with the underlying features proposed in this paper to achieve super-resolution reconstruction,and then uses the convolutional blind denoising network for denoising.The effectiveness of the integrated network for low-quality image super-resolution reconstruction is verified by experiments,supplemented by denoising and post-processing,which further improves the quality of infrastructure visual monitoring images.