Research On The Railway Low Illumination Image Enhancement Based On The Improved AOD-Net

In recent years,computer vision technology is widely used in railway safety detection,which has the advantages of high accuracy and non-contact detection.However,in the low illumination environment,such as tunnel and night,the quality of the collected images is seriously degraded,which leads to the difficulty of feature extraction and limits the effectiveness of the computer vision system.Therefore,to improve the reliability of the computer vision system in the railway illumination environment and realize the all-weather work of the railway detection system,it is necessary to study the railway low illumination image enhancement technology.Firstly,this thesis analyzes the existing railway low illumination image enhancement methods.It is pointed out that the existing methods have two disadvantages.The first disadvantage is using the dark channel prior dehazing algorithm to enhance the railway low illumination images,resulting in poor real-time performance.The second disadvantage is using a threshold to segment the highlighted areas in the low illumination images,the fixed threshold results in the limited application of the algorithms.Given the problems of the existing algorithms,combined with the characteristics of the railway low illumination images,a railway low illumination image enhancement method based on the improved All-in-One Dehazing Network(AOD-Net)is proposed.Secondly,this thesis analyzes AOD-Net and points out the disadvantages of AOD-Net.And this thesis makes improvements.On the one hand,aiming at the problem that the receptive field of AOD-Net is too small to extract large-scale features,which may cause the halo effect after dehazing,this thesis introduces hybrid dilated convolution to expand the receptive field.The hybrid dilated convolution can expand the receptive field without increasing the cost of computing and reducing the resolution of feature maps.On the other hand,aiming at the problem that AOD-Net treats different levels of features equally,the feature fusion module is designed to give different weights to different levels of features.To verify the effectiveness of the improved AOD-Net,a comparative experiment is carried out on the synthetic haze images and the real haze images.The experimental results show that the improved AOD-Net has better performance in dehazing,and its real-time performance is far better than the dark channel prior dehazing algorithm.Finally,the improved AOD-Net is used to enhance the railway low illumination images.To solve the halo effect in the highlighted areas of the low illumination images,using a series of gamma transform to enhance the low illumination images.The images enhanced by gamma transform and the image enhanced by the improved AOD-Net are fused by image fusion technology to eliminate the halo effect and realize the enhancement of the whole low illumination image.To verify the effectiveness of our algorithm,a comparative test is carried out on the railway low illumination images.The result shows that the algorithm in this thesis is better than the existing algorithms in contrast enhancement,detail information recovery,operation time,and application scope.