Noise Modeling and Denoising of Images Collected by On-Board Track Inspection System

With the increase of railway mileage year by year, technologies such as vision and optical-fiber sensing are widely applied to automatic railway inspections. On-board track inspection system (OBTIS) is a vision-based system designed for automatic railway inspection. How to improve the defect detection accuracy of the images collected by the OBTIS is a practical problem. Images will inevitably be contaminated by noises in the process of acquisition. And noise will have a negative effect on the accuracy of defect detection. To reduce the influence of noise on defect detection, on the one hand, we model the noise in the images collected by OBTIS as region-adaptive (RA) Gaussian noise through the probability distribution characteristics of numerous images. On the other hand, based on weighted nuclear norm minimization (WNNM), we propose a denoising model named RA-WNNM to reduce the noise of images in OBTIS. RA-WNNM adopts the adaptation characteristic of noise into a weight matrix, which makes it have no analytical solution. The alternating direction method of multipliers framework is employed to decompose the RA-WNNM into several sub-problems with analytical solutions for iterative solution. Experimental results demonstrate the superiority of RA-WNNM both in denoising and image classification.