Inspeção de Waysides usando Processamento de Imagens e Deep Learning

Abstract

The advances on deep learning techniques made computer vision tasks more accurate and faster by relying on convolutional neural networks and more powerful hardware. In the industry, automatic inspection supported by these methods may ensure constant maintenance and avoid railway acidentes. Thereby, this work proposes the application of deep learning and image processing methods for automatic inspection of train wagons wheelsets. More specifically, the size of wheels and the thickness of bandage are measured, in addition to locating the bearing fixing screws. The neural network built performs semantic segmentation on photographs provided by the mining company Vale. Using a U-Net architecture with ResNet50 as backbone, the network was able to reach 92.50% in mIoU and 97.52% in mPA, which are the adopted metrics for evaluating this proposal. The post-processing step retrieved the screws and improved evaluation metrics, indicating the success of the proposed inspection.