Efficient instance segmentation framework for UAV-based pavement distress detection

Pavement distress detection is critical for ensuring road safety. Recently, Unmanned Aerial Vehicles (UAVs) become an efficient means of capturing large-scale pavement images. However, traditional pavement distress detection methods face challenges with UAV images: object detection lacks pixel-level information, while semantic segmentation fails to differentiate between individual instances. This paper introduces PDIS-Net, an instance segmentation framework specifically designed for UAV-based pavement distress detection. PDIS-Net first employs a fully dynamic convolution kernel generation strategy, predicting both kernel positions and weights. These kernels are then optimized via metric learning and kernel fusion. Finally, these high-quality kernels are convolved with feature maps to produce accurate instance masks. Experimental results on the UAPD-Instance dataset reveal that PDIS-Net achieves a mean average precision (mAP) of 78.1% at 30.8 FPS, outperforming other methods by 15.4%. Furthermore, real-world tests validate the robustness and effectiveness of PDIS-Net in highway pavement distress detection, highlighting its potential for practical deployment.