Noise Reduction of Rail Surface Defect Images Based on Attention-guided Poly-scale Denoising Convolutional Neural Networks

In order to solve the problem of the dependence of noise reduction of rail surface defect images on manual setting of filtering parameters and blurring of defect edges, an attention-guided poly-scale noise reduction convolutional neural network-based noise reduction method was proposed for rail surface defect images. Firstly, the poly-scale convolution in the deep network was used to automatically extract the features of noise-containing images, to avoid relying on manually set filtering parameters and overcome the problem of blurred defect edges caused by insufficient refinement of single-scale convolutional features. Secondly, the deep and shallow features of the network were fused using jumping connections to strengthen the influence of the shallow features and overcome the problem of the shallow features being ignored in the deep layer due to the deeper network, to deliver more adequate features. Thirdly, attention mechanism was used to adjust the weights of features at different locations in space to filter out features that can characterize noise, and obtain noise information. Finally, the noise information in noisy images was removed by the reconstruction module to achieve end-to-end noise reduction. The experimental results demonstrate qualitatively and quantitatively that the proposed method is more effective both in noise reduction and in retaining defect edge information, providing conditions for accurate defect segmentation. © 2024 Science Press. All rights reserved.