BD-YOLOv8s: enhancing bridge defect detection with multidimensional attention and precision reconstruction

Abstract

The YOLO (You Only Look Once) series has recently demonstrated remarkable effectiveness in the domain of object detection. However, deploying these networks for concrete bridge defect detection presents multiple challenges, such as insufficient accuracy, missed detections, and false positives. These complications arise chiefly from the complex backgrounds and the significant variability in defect characteristics observed in bridge imagery. This study presents BD-YOLOv8s, an advanced methodology utilizing YOLOv8s for bridge defect detection. This approach augments the network’s adaptability to a broad spectrum of bridge defect images through the integration of ODConv into the second convolutional layer, processing information within a four-dimensional kernel space. Furthermore, incorporating the CBAM module into the first two C2F architectures leverages spatial and channel attention mechanisms to focus on critical features, thus enhancing the accuracy of detail detection. CARAFE replaces traditional upsampling methods, improving feature map reconstruction and significantly reducing blurs and artifacts. In performance assessments, BD-YOLOv8s attained 86.2% mAP@0.5 and 56% mAP@0.5:0.95, surpassing the baseline by 5.3% and 5.7%. This signifies a considerable decrease in both false positives and missed detections, culminating in an overall improvement in accuracy.