Lightweight SM-YOLOv5 Tomato Fruit Detection Algorithm for Plant Factory

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Robots and intelligent devices are crucial for modern precision agriculture.
• Accurate tomato detection is essential for automated systems in plant factories.
• Existing methods struggle with small-target detection accuracy due to environmental complexity and computational limits.

Purpose of the Study:

• To develop a lightweight and efficient algorithm for small tomato detection in plant factories.
• To improve the precision and recall of target detection for agricultural robots.
• To address the limitations of current computer vision models in complex agricultural environments.

Main Methods:

• An improved Small MobileNet YOLOv5 (SM-YOLOv5) detection algorithm was proposed.
• MobileNetV3-Large was integrated as a lightweight backbone network.
• A dedicated small-target detection layer was incorporated into the YOLOv5 architecture.
• The model was trained on a custom plant factory tomato dataset.

Main Results:

• The SM-YOLOv5 model achieved a mean Average Precision (mAP) of 98.8%, a 1.4% increase over the baseline YOLOv5.
• Model size was reduced by 57.52% (to 6.33 MB), and computational requirements (GFLOPs) were halved.
• Achieved a precision of 97.8% and a recall rate of 96.7%.

Conclusions:

• The SM-YOLOv5 model offers a lightweight and high-performance solution for small tomato detection.
• The improved model meets the real-time detection demands of tomato-picking robots in plant factories.
• This advancement supports the development of more efficient and intelligent agricultural automation.