Railway Wheel-Tread Defect-Recognition Method Using Improved Convolutional Neural Network Technology

Wheel-tread defect recognition is a crucial step in ensuring the safety of train wheel-rail system services. However, the diverse and complex nature of wheel-tread defects, coupled with the presence of minor defect features, poses significant challenges in accurately identifying defects using existing deep convolutional neural networks. To address this problem, we developed a small target defect-detection module and proposed a railway wheel-tread defect-recognition method based on an improved convolutional neural network. First, a deformable convolutional attention-enhanced bottleneck module was designed to achieve an adaptive adjustment of the network receptive field in the backbone network. Second, an adaptive spatial and channel enhancement module was constructed to further improve the network’s sensitivity and processing capabilities for different features. Third, we designed a new module called the spatial grouped attention fusion pyramid module to enhance the extraction and fusion capabilities of multiscale features through grouping and fusion of spatial attention mechanisms, enabling the effective extraction and discrimination of defect multi-layer semantic features. Finally, experiments were conducted using a tread defect dataset with an imbalance ratio of 10:1. The experimental results demonstrated the excellent performance of the proposed model on public datasets. The average mAP@0.5 value increased from 90.6% to 91.9%. Similarly, the observed average mAP@0.5:0.95 value increased from 53.9% to 54.9%.