Detection of road defects with weak small samples based on multiple deep learning models

Potholes on roads not only jeopardize traffic safety and driving comfort but also necessitate efficient detection and maintenance measures. Traditional manual detection methods are labor-intensive and time-consuming. Thus, developing automated solutions is imperative. This study addresses this challenge by constructing a dataset and evaluating various deep learning models, including GoogleNet, VGG16, ResNet50, AlexNet, YOLOV5, and YOLOV8, for pothole detection. YOLOV8 emerges as the optimal choice due to its superior accuracy. However, accurately estimating pothole areas proves challenging due to their irregular shapes. To mitigate this, an innovative algorithm is proposed, integrating YOLO with a pre-trained segmentation model. This enables precise pixel-level delineation of pothole areas. Additionally, the algorithm incorporates the intersection over union (IOU) metric to calculate the ratio of pothole area to the total image area. By enhancing both detection and area estimation, this approach holds promise for improving road safety and facilitating maintenance efforts. Automated detection and accurate area estimation not only save time and resources but also provide crucial data for prioritizing and planning road repair and maintenance tasks.