Research On Wheelset Laser Stripe Image Inpainting Based On Deep Learning

The railway is an important transportation facility in China and the main artery of China's economic development.Wheelset is a key moving component of locomotive,and its fault seriously threatens the operation safety of railway and urban rail transit.Therefore,it is very important to develop real-time,high-precision and adaptive wheelset intelligent monitoring equipment.Since wheelset light stripe images are collected outdoors dynamically and affected by environmental interference,the images contain light spots and local fractures.Therefore,it is necessary to build an image inpainting model to remove the light spots,complement the fractures and restore the light stripe textures in the images.At present,the deep-learning-based models have a remarkable effect on image restoration on RGB public datasets such as faces and outdoor scenes.However,the effect is not satisfying on wheelset laser stripe images(grayscale images),especially in large spot regions and local fracture regions.In addition,since the inpainting model runs on the embedded monitoring equipment,it should be both lightweight and high precision.The main contributions of this paper are as below:(1)An image inpainting model RSM-Net(Recurrent Similarity Mapping Network)based on recurrent neural network is proposed for wheelset image inpainting.The model replaces the original Pconv layer with soft-coding Pconv(Partial convolution)layer to strengthen the feature learning ability and improve the accuracy effectively.The asymmetrical similarity module is designed to calculate both the angle difference between the feature vectors and the response value of the source feature vectors.Thus,the light stripe features can be applied to optimize the unknown region inpainting more effectively.The multi-scale structural similarity(MS-SSIM)loss term is introduced into the mixed loss function(the absolute difference term,the perceptual loss term,the style loss term and the total variation loss term)to precisely guide the structural information restoration of the wheelset light stripes and realize high-precision light stripe image restoration.(2)The contrast experiments are conducted on RSM-Net and other mainstream image inpainting models(Pconv-Net,LBAM,RFR-Net,etc.).The experiments demonstrate that the proposed RSM-Net has the state-of-the-art accuracy in the small spot regions,the large spot regions and the fracture regions of the wheelset laser stripe images.(3)LRSM-Net(Lightweight Recurrent Similarity Mapping Network)is proposed by optimizing RSM-Net.LRSM-Net applies linear mapping convolution,which can retain the core features of wheelset images as much as possible,remove redundant features and expand the channel features through the linear transformation of the core features,and then realize lightweight.(4)The contrast experiments are conducted on RSM-Net,LRSM-Net and the lightweight model with depthwise separable convolution.The experiments demonstrate that the lightweight model with depthwise separable convolution can reduce the parameters by about 80% compared with RSM-Net,but the inpainting accuracy is lost seriously.Compared with RSM-Net,LRSM-Net can reduce the parameters by nearly50%,and the loss of inpainting precision is limited.Combined,the proposed LRSM-Net is comprehensive optimal.The experiments demonstrate that the RSM-Net proposed in this paper is superior to other mainstream models in the task of image restoration for wheelset laser stripe images.The LRSM-Net proposed can achieve comprehensive optimal in lightweight and high-precison,and meet the requirements of wheelset monitoring equipment.