Road Surface Defect Detection Based on Partial Convolution and Global Attention

The balance between detection accuracy and real-time performance in road surface damage detection models has not been achieved. To address the characteristics of complex backgrounds, significant variations in target shapes and scales, a large number of small targets, and high real-time requirements in road surface damage images, we propose the FG_YOLO object detection algorithm based on YOLOv5s. To effectively reduce model parameters and improve detection speed, we designed the FasterBottle Neck module based on Partial Convolution and Separable Convolution to replace the main computational modules in YOLOv5s. This approach reduces model parameters while effectively enhancing detection accuracy. Additionally, we incorporated a Global Attention Mechanism during the feature extraction stage to significantly improve the model’s ability to extract features for targets with varying shapes and small targets in complex backgrounds, thereby enhancing the detection capability for small targets. To address the imbalance between positive and negative labels during the training phase, we replaced the loss function to accelerate model convergence. Validation on the public dataset GRDDC2020 shows that the algorithm achieves an average precision of 55.8%, a detection speed of 115 frames per second, and a model parameter size of 6.08 M. Compared to YOLOv5s, the average precision increased by 4.7%, the detection speed increased by 3 frames per second, and the model parameters decreased by 13.4%, meeting the requirements for both detection accuracy and speed in road surface damage detection. The improved model’s effectiveness is also demonstrated on the UAV-PDD2023 dataset.