An Ensemble Learning Approach With Attention Mechanism for Detecting Pavement Distress and Disaster-Induced Road Damage
Shouxing Wang, Hongzan Jiao, Xin Su, Qiangqiang Yuan
Abstract
Road damage presents a significant risk to traffic safety, including pavement distress and disaster-induced damage. Thanks to their high efficiency, computer vision-based methods for pavement distress detection have been widely developed. In disaster scenarios, the automatic extraction of road damage information from extensive social media images plays a critical role in rescue efforts. However, few existing studies have focused on detecting object-level disaster-induced road damage. To fill the gap, this paper presents a Social media image dataset of Object detection for Disaster-induced Road damage (SODR), including 1,552 images and two categories (i.e., collapses and blockages). Additionally, this paper proposes an ensemble learning approach with attention mechanisms based on YOLOv5 (You Only Look Once) network. Initially, attention modules are employed to create two distinct detectors for ensemble learning. Subsequently, one standard YOLOv5 and two variant networks are trained with consistent settings, and test time augmentation is applied during the inference phase. The proposed method has been implemented across five scales of YOLOv5, offering alternatives for balancing accuracy and computational cost. To demonstrate the validity, comprehensive experiments were conducted on two datasets. Compared with some mainstream detectors and ensemble learning methods, our approach achieved competitive results with a fewer number of parameters and a simpler training and testing process. The SODR dataset and source code are available at (https://github.com/nonondayo/yolov5_SODRv1).