Abstract:

In order to address the issue of missed detections during small target detection of aircraft rivets, an improved YOLOv8n algorithm for the detection of aircraft rivets and their anomalies was proposed. First, by adding a small object detection head, the shallow detail information in the backbone network was better fused, enhancing the model's feature fusion capability and its ability to recognize and locate small rivet targets. Second, the first two convolutions in the backbone network were replaced with SPD-Conv, which reduces information loss during down sampling through the combination of feature map reorganization and non-stride convolutions. Finally, large separable kernel attention (LSKA) was integrated into the spatial pyramid pooling fast (SPPF) module, capturing the dependencies between spatial and channel dimensions by calculating spatial and channel weights on each feature map and adjusting the feature maps to enhance the algorithm's ability to extract and recognize rivet feature information. Ablation experiments and comparative experiments were conducted based on a self-built aircraft rivet dataset. The results show that the proposed algorithm can achieve real-time identification of aircraft rivets and their anomalies, with precision, recall, and mean average precision (mAP) values improved by 6.5%, 16%, and 15%, respectively, compared to the YOLOv8n algorithm. The detection performance is also significantly better than other mainstream algorithms.

Key words: improved YOLOv8n, aircraft rivets, detachment, anomaly detection, space to depth convolution(SPD-Conv), ablation experiment