YOLO-MoE: Multi-Scenario Cigarette Defect Detection with a Mixture of Experts

Cigarette defect detection plays a crucial role in quality control during cigarette production. Existing defect detection methods are only applicable to single scenarios, making it difficult to integrate and unify detection systems. To address the demand for multi-scenario cigarette defect detection, this paper proposes YOLO-MoE, a mixture-of-experts-based object detection method built upon YOLOv5. The C3 is enhanced with a routing network and a multi-expert network: the routing network adaptively assigns dynamic expert weights to different input features, while the multi-expert network extracts expert features in parallel and aggregates them through weighted fusion based on expert weights. In addition, to mitigate the issue of chaotic weight allocation, an expert weight assignment loss is proposed to guide the routing network in assigning weights more accurately. Experiments conducted on datasets constructed from four scenarios, including 1 cigarette stick and 3 cigarette packs, demonstrate that YOLO-MoE outperforms baseline models, validating its effectiveness in enhancing the practical application of cigarette defect detection.