Aircraft skin damage detection method integrating multi-scale feature reconstruction and adaptive loss weight

Aircraft skin damage poses a significant threat to the safety and airworthiness of aircraft. The considerable variation in the size of damage objects introduces challenges in feature extraction and compromises the accuracy of detection models. To address these issues, this paper proposes an improved aircraft skin damage detection algorithm based on YOLOv9s. In the feature extraction stage, a lightweight Inception module is introduced to enhance the representation of multi-scale features while reducing the model's parameter count. In the feature fusion stage, the SKFusion module, incorporating the SKNet attention mechanism, is introduced to further improve the model's ability to capture damage features at different scales. Additionally, a dynamic weighted optimization strategy combining an improved Focaler loss and inner-DIoU loss function is employed to enhance the recognition of hard-to-classify samples. The proposed method was evaluated on both a self-constructed aircraft skin damage dataset and the publicly available NEU-DET dataset. Experimental results demonstrate that, compared with the original YOLOv9s, the improved model achieves a 1.6% and 3.9% increase in mAP@50 on the self-constructed and NEU-DET datasets, respectively, while reducing the model's parameters by 16.7%.