Lightweight tomato ripeness detection algorithm based on the improved RT-DETR.

Area of Science:

• Computer Vision and Machine Learning
• Agricultural Robotics
• Horticultural Science

Background:

• Accurate tomato ripeness identification is crucial for efficient harvesting and economic benefits in agricultural management.
• Existing intelligent harvesting systems lack fine-grained ripeness detection, leading to losses from harvesting immature or rotten fruits.
• Environmental factors like uneven lighting and fruit occlusion challenge robotic ripeness assessment, necessitating robust and efficient models.

Purpose of the Study:

• To develop a lightweight and accurate model for fine-grained tomato ripeness detection in challenging natural conditions.
• To enhance the efficiency and economic viability of intelligent tomato harvesting systems.
• To address limitations of existing models in terms of accuracy, speed, and computational cost.

Main Methods:

• Proposed a novel lightweight model, PDSI-RTDETR, incorporating a PConv_Block module for efficient feature extraction.
• Integrated a deformable attention module with intra-scale feature interaction for detailed feature analysis.
• Developed a slimneck-SSFF feature fusion structure and an Inner-EIoU loss function to optimize computation and detection accuracy.

Main Results:

• PDSI-RTDETR achieved an mAP50 of 86.8%, a 3.9% improvement over the baseline RT-DETR.
• The model demonstrated a 38.7% increase in frames per second (FPS) and a 17.6% reduction in GFLOPs.
• The model outperformed existing methods in precision and speed, showing significant potential for real-world applications.

Conclusions:

• The PDSI-RTDETR model offers a highly accurate and efficient solution for tomato ripeness detection.
• Its lightweight design and improved performance make it suitable for deployment on intelligent harvesting robots.
• This advancement can significantly enhance tomato harvesting quality by minimizing the collection of undesirable fruits.