LMS-Res-YOLO: Lightweight and Multi-Scale Cucumber Detection Model with Residual Blocks

Efficient cucumber detection in greenhouse environments is crucial for agricultural automation, yet challenges like background interference, target occlusion, and resource constraints of edge devices hinder existing solutions. This paper proposes LMS-Res-YOLO, a lightweight multi-scale cucumber detection model with three key innovations: (1) A plug-and-play HEU module (High-Efficiency Unit with residual blocks) that enhances multi-scale feature representation while reducing computational redundancy. (2) A DE-HEAD (Decoupled and Efficient detection HEAD) that reduces the number of model parameters, floating-point operations (FLOPs), and model size. (3) Integration of KernelWarehouse dynamic convolution (KWConv) to balance parameter efficiency and feature expression. Experimental results demonstrate that our model achieves 97.9% mAP@0.5 (0.7% improvement over benchmark model YOLOv8_n), 87.8% mAP@0.5:0.95 (2.3% improvement), and a 95.9% F1-score (0.7% improvement), while reducing FLOPs by 33.3% and parameters by 19.3%. The model shows superior performance in challenging cucumber detection scenarios, with potential applications in edge devices.