Sugarcane stem node identification algorithm based on improved YOLOv5.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Sugarcane stem node identification for seed pre-cutting machines typically requires high-performance equipment, leading to inefficiencies.
• Existing methods lack the optimal balance between accuracy and computational efficiency for practical agricultural applications.

Purpose of the Study:

• To develop a novel, lightweight deep learning architecture for efficient and accurate sugarcane stem node detection.
• To optimize the YOLOv5 framework for enhanced performance in agricultural machinery.

Main Methods:

• Proposed a novel lightweight architecture, G-YOLOv5s-SS, based on YOLOv5.
• Modified the backbone and neck structures to utilize shallow-level features and reduce complexity.
• Integrated Ghost lightweight modules and the SimAM attention mechanism to improve feature extraction and accuracy.

Main Results:

• G-YOLOv5s-SS achieved an average precision of 97.6% for sugarcane stem node identification, outperforming the YOLOv5 baseline by 0.6%.
• The model size was reduced to 2.6MB, with 1,129,340 parameters and 7.2G FLOPs, representing significant reductions in size and complexity.
• Demonstrated superior performance compared to various mainstream one-stage object detection algorithms in both accuracy and model size.

Conclusions:

• The G-YOLOv5s-SS architecture offers a highly accurate and computationally efficient solution for sugarcane stem node detection.
• The developed model meets the performance requirements for integration into sugarcane seed pre-cutting machines.
• This lightweight approach balances high recognition accuracy with reduced model complexity, making it suitable for resource-constrained environments.