D-Triple: An Optimized Defect Detection Deep Model for Sewer Pipes

Sewer pipes are an fundamental infrastructure of modern cities, and their automatic maintenance is an important requirement for the operation of smart cities. The key issue for sewer pipe maintenance is the detection of their defects, which can lead to various levels of maintenance requirements. In this work, we present an optimized deep learning model for defect detection of sewer pipes. The model is optimized in three aspects: the backbone, the neck, and the detection head. The backbone is replaced with Swin Transformer, the neck is enhanced with Feature Pyramid Network (FPN) and RoI align, and the detection head is optimized by modifying the loss function with GIoU and introducing a new hyperparameter on classification loss. With these optimizations, the model has better feature extraction capability, better small defect detection, and better loss function for model training. Combined with data augmentation techniques, the optimized model achieves significant improvements over the baseline Faster R-CNN model, with mAP increased from 0.606 to 0.743. In addition to mAP improvements, our optimized model is especially effective on finding some sewer defects that are difficult to recognize.