YOLO-MCD: An Efficient Detection Method for Multi-Dish Recognition

In response to the challenges of insufficient detection accuracy, limited processing speed, and poor recognition of long-tail categories in intelligent catering scenarios, this paper proposes YOLO-MCD (YOLO-multi-Chinese-dish), a fast and accurate multi-dish detection method. Based on YOLO11, YOLO-MCD introduces three key improvements: the design of the C3K2_RFCBAMConv module to enhance feature extraction capability; the incorporation of an EUCB upsampling structure to strengthen multi-scale feature fusion; and the construction of a dynamic loss function EMASlideLoss, which integrates the exponential moving average with an IoU-based adaptive weighting mechanism to significantly improve the detection performance of long-tail categories. Experiments on a self-built multi-dish dataset demonstrate that YOLO-MCD achieves 91.2% mAP@50 and 86.4% mAP@50-95, with an inference speed of 278 FPS, outperforming YOLO11n by 2.2% in mAP@50 while increasing the parameter count by only 7%. Compared with SSD, Faster R-CNN, and RT-DETR-R18, YOLO-MCD exhibits superior performance in both accuracy and speed. Furthermore, ablation studies verify the independent effectiveness and synergistic gains of each module. YOLO-MCD maintains real-time inference capability while achieving high-precision detection for long-tail distributions, multi-object scenes, and complex backgrounds, demonstrating great potential for applications in intelligent catering and food safety.