CSM-YOLO:A Lightweight and High-Precision Network Geared to Defects on Aircraft Surface in Detecting

JIE Zhanduo; ZHANG Zhengming; HUANG Haoran; HAO Ming; ZHAO Jianbang

doi:10.3969/j.issn.2097-1915.2025.05.002

Abstract：In view of the problems that in detecting defects often show on surface of aircraft by using exist-ing visual-based method,there remains a low accuracy in detection,high parameters and expend time is more in computational consumption,and such a method is difficult to balance accuracy improvement with model lightweight,a new high-precision and lightweight aircraft surface defect detection method,called CSM-YOLO,is proposed.First,C2f module in the backbone network replaced by C2f-SCSA module is to dynamically enhance multi-scale features and improve the model's ability to capture,extract,and utilize key feature information,solving the feature loss caused by down-sampling.Secondly,the Slim-Neck fea-ture fusion network is improved with cross-layer connection and applied to the model neck,realizing the boosted computational efficiency and the detection accuracy and simultaneously cutting the information loss.Lastly,MPDIoU Loss is used to enhance bounding box regression accuracy and improve the detection precision of small target defects and cut false and missed detections.The experiments show that the CSM-YOLO enables to achieve high precision and lightweight.A maximum detection accuracy of 88.34%on surface defects can be reached,and there is a 2.92%improvement compared with the baseline YOLOv8n model,and the improvement in computation is obvious compared with the YOLOv3-tiny,the YOLOv5n,the YOLOv5s,the YOLOv7-tiny,the YOLOv9t and the YOLOv12n respectively.In aspect of model pa-rameters and computation,for the CSM-YOLO,there is a parameter of 2.67×106/s and a computational cost of 7.68×109/s,reducing the baseline YOLOv8n by 0.34×106/s and 0.41×109/s respectively.And the CSM-YOLO is balancing the accuracy improvement with the model lightweight.Moreover,the CSM-YOLO delivered significant performance gains on the aircraft surface defect detection dataset,offering an effective automated solution for surface defect detection.