Tomato ripeness and stem recognition based on improved YOLOX.

Area of Science:

• Computer Vision
• Agricultural Robotics
• Machine Learning

Background:

• Intelligent harvesting faces challenges with unbalanced tomato fruit maturity levels and low recognition accuracy for fruits and stems.
• Existing models struggle with scale variations and class imbalance in agricultural applications.

Purpose of the Study:

• To develop an improved object detection model for accurate identification of tomato fruit maturity and stems in intelligent harvesting.
• To address class imbalance and scale variations in tomato detection datasets.

Main Methods:

• Proposed the YOLOX-SE-GIoU model, integrating an SE focus module into YOLOX for improved identification accuracy.
• Optimized the loss function to GIoU loss to handle scale discrepancies in fruits and stems.
• Evaluated the model against YOLOv4, YOLOv5, YOLOv7, and the original YOLOX.

Main Results:

• The YOLOX-SE-GIoU model achieved a mean average precision (mAP) of 92.17%.
• Demonstrated significant improvements in average precision (AP) for semi-ripe tomatoes (1.68–26.66% increase) and stems (3.78–45.03% increase).
• Outperformed existing models, showing 1.17–22.21% higher mAP compared to YOLOv4, YOLOv5, YOLOv7, and YOLOX.

Conclusions:

• The YOLOX-SE-GIoU model offers superior recognition performance for unbalanced and scale-variant tomato samples.
• Effectively reduces false and missed detections, enhancing accuracy in automated tomato harvesting.
• Provides a foundational technology for advancing fruit harvesting automation.