Abstract Accurate detection of metal roof weld seam defects is crucial for ensuring the structural safety of buildings. Traditional detection methods often lack sufficient sensitivity when dealing with complex weld seam surface textures, dynamic lighting conditions, and the identification of subtle defects. To address these industrial challenges, this paper proposes a lightweight detection algorithm—CBI-YOLO, which innovatively integrates Cross-Stage Partial Path Convolution (CSPPC), Bi-directional Feature Pyramid Network (BiFPN), and Inner Intersection over Union (InnerIoU). Based on the YOLOv8n backbone network, the proposed method achieves efficient real-time performance. Experimental results show that CBI-YOLO achieves precision, recall, and F1-score of 0.873, 0.843, and 0.877, respectively, with the model's parameter count reduced from 3,006,428 to 1,933,224, GFLOPs decreased from 8.1 to 5.9, and the model size compressed from 6.3MB to 4.1MB. Through structural optimization, the proposed method reduces parameters, computational complexity, and model size by approximately 35%, 27%, and 35%, respectively, significantly improving efficiency and compression performance. Extensive experiments conducted on a dataset containing various types of metal roof weld seam defects demonstrate that CBI-YOLO maintains robust detection performance even under complex lighting conditions. This research provides an intelligent solution with both high precision and efficiency for metal roof structural health monitoring.
Bai et al. (Fri,) studied this question.