Detection Algorithm for Air Duct Clamp on Trains Based on RSA-YOLOv10n

To address challenges in the detection of air duct clamps due to factors such as dim environments and small objects in train detection, such as insufficient detection accuracy and low detection efficiency, this paper proposes a RSA-YOLOv10n-based detection algorithm, developed utilizing a dataset of air duct clamps collected from trains. Firstly, in order to better capture diverse features in images, the Conv convolution is modified to RepConv convolution based on the YOLOv10n model, facilitating adaptive adjustments in the representation ability of the convolution kernel. Secondly, in order to improve the sensitivity of the model to important information, the StokenAttention attention mechanism is introduced at the Neck layer, supporting dynamic focus on key areas during image detection. Experiments conducted using a dataset comprising thousands of self-labeled images of air duct clamps demonstrated that the <italic>mAP</italic>@50%, <italic>mAP</italic>@95%, accuracy, and recall rate of the proposed RSA-YOLOv10n model were 98.7%, 75.1%, 97.3%, and 96.6%, respectively. These results show improvements of 0.2, 0.8, 0.2, and 0.3 percentage points compared to the YOLOv10n model. Furthermore, the RSA-YOLOv10n model proves a great improvement in the FPS value, while maintaining model size, GFLOPs, function loss value, and other parameters equivalent to those of the YOLOv10n model. Through ablation experiments and comparative experiments with other models, including YOLOv3-tiny, YOLOv5n, and YOLOv8n, the analysis and verification showcased the superiority of the proposed detection algorithm in terms of overall accuracy, recall rate and detection time, suggesting its potential for widespread application in the field of detecting air duct clamps on trains. The study findings are of great significance for the technical development of train device and component detection.