An Improved YOLOv8 Detection Model for Catenary Components Using Long-Distance Feature Dependence

In the condition monitoring system of high-speed railway catenary support components, the positioning and recognition performance directly affects the performance of the state detection tasks such as anomaly detection and defect recognition. Due to limitations such as complex background and long-distance feature extraction, traditional defect detection methods have difficulty fully exerting their detection performance. Therefore, an improved YOLOv8 model is proposed to solve these detection problems. First, a long short-term memory (LSTM) module is added to the backbone middle layer to more effectively capture the object's long-distance dependencies, improving the ability to extract long-distance features in sequence data. Secondly, a sizeable separable kernel attention (LSKA) module is introduced to the spatial pyramid pooling feature (SPPF) layer to further improve the model's ability to capture long-range image dependencies. Experimental results show that the detection framework achieves a detection accuracy (mean average precision, mAP) of 75.3% while maintaining low computational complexity, proving its effectiveness in catenary detection. Therefore, the proposed method can be effectively applied to the detection task of catenary components.