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

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Efficient cucumber detection is vital for automated greenhouse farming.
• Existing methods face challenges with background interference, occlusion, and edge device limitations.

Purpose of the Study:

• To develop a lightweight, multi-scale cucumber detection model (LMS-Res-YOLO) for enhanced agricultural automation.
• To address computational constraints of edge devices while maintaining high detection accuracy.

Main Methods:

• Proposed LMS-Res-YOLO model incorporating a High-Efficiency Unit (HEU) with residual blocks.
• Implemented a Decoupled and Efficient detection HEAD (DE-HEAD) to reduce model complexity.
• Integrated KernelWarehouse dynamic convolution (KWConv) for optimized parameter efficiency and feature expression.

Main Results:

• Achieved 97.9% mAP@0.5 and 87.8% mAP@0.5:0.95, outperforming the benchmark YOLOv8_n.
• Reduced Floating-Point Operations (FLOPs) by 33.3% and model parameters by 19.3%.
• Attained a 95.9% F1-score, demonstrating robust performance in challenging detection scenarios.

Conclusions:

• LMS-Res-YOLO offers a computationally efficient and accurate solution for cucumber detection in greenhouses.
• The model's lightweight design makes it suitable for deployment on edge devices, advancing agricultural automation.
• Innovations like HEU, DE-HEAD, and KWConv contribute to superior performance and efficiency.