Automatic Rail Surface Defect Inspection Using the Pixelwise Semantic Segmentation Model

Abstract

Rail surface defect (RSD) is an important railroad track quality indicator and impacts the overall track safety and ride quality. Railroads need to inspect and evaluate RSD conditions regularly to make proper maintenance plan for both freight and passenger services. Unfortunately, current field practices for rail surface defect defection have not been fully automated yet and significant amount of manual inspections are still involved, which could be labor-intensive but low-efficient. Earlier efforts have proposed multiple automatic rail surface defect detection systems, but the accuracy and equipment cost issues have limited their applications in the field. To address the needs in track inspections, this study proposes an improved Deeplabv3-plus model using a lightweight backbone, attention module, and the Lovász-Softmax loss to automatically detect RSD. The improved model not only improves the detection performance but also keeps the computational cost manageable. The ResNet-18 backbone is adopted for the encoder design to reduce the computational cost and maximize the inference speed. The CBAM attention module is unified with the decoder structure for critical features’ representation. The optimum crop size is determined in the training process. The Lovász-Softmax loss is implemented to address the severe class imbalance issue. The improved DeepLabv3-plus model is compared with five other models to validate the performance. Experimental results on both original data and noisy data confirm the proposed model achieves the best performance on evaluation metrics and visualization. This work provides a feasible solution for the future implementation of automatic railroad track inspection.