MLM-YOLO: lightweight improved YOLOv11n model for vehicle-mounted roadside garbage bin detection

Abstract

Vehicle-mounted roadside garbage bin detection faces several challenges, including large target-scale variations, complex background interference, similar appearances among categories, and limited computational resources for edge deployment, which may lead to missed detections, false detections, category confusion, and reduced deployment efficiency. To address these problems, this paper proposes MLM-YOLO, a lightweight detection model based on YOLOv11n, for recognizing and detecting four types of garbage bins. First, C3k2-MDSF is adopted in the backbone to replace part of the original C3k2 modules, enhancing the multi-scale representation of local texture, edge contours, and spatial structure information. Second, C2PSA-MFDSA is introduced at the end of the backbone to replace the original C2PSA module, strengthening discriminative high-level semantic features while suppressing complex background noise. Finally, lightweight dynamic sampling upsampling is employed in the neck upsampling stage to replace fixed interpolation, improving the adaptive reconstruction of high-level semantic features. Experiments on a self-built garbage bin dataset show that MLM-YOLO improves Precision, Recall, mAP@0.5, and mAP@0.5:0.95 by 3.2, 0.6, 1.1, and 1.6 percentage points, respectively, compared with YOLOv11n, while reducing parameters by 5.0% and computational cost by 1.6%. Although the inference speed of MLM-YOLO decreases from 72.96 FPS to 63.22 FPS under the desktop evaluation setting, it still maintains real-time inference capability. Furthermore, edge-device deployment evaluation on the Jetson Orin Nano shows that MLM-YOLO achieves 36.2 FPS, indicating its inference-level deployment feasibility for vehicle-mounted roadside garbage-bin detection.