TRD-Net: an efficient tomato ripeness detection network based on improved YOLO v8 for selective harvesting

Area of Science:

• Agricultural Engineering
• Computer Vision
• Machine Learning

Background:

• Selective harvesting requires accurate fruit recognition and ripeness detection.
• Existing methods face challenges in complex environments with occlusions and lighting variations.

Purpose of the Study:

• To develop a novel, lightweight tomato ripeness detection network (TRD-Net) for improved performance in real-world conditions.
• To enhance real-time tomato recognition and maturity detection for automated harvesting systems.

Main Methods:

• A tomato dataset of 3,330 real-world images was created.
• The YOLO v8s network was improved using spatial and channel reconstruction convolution (SCRConv) and SimAM attention.
• The MPDIoU loss function replaced CIoU loss for enhanced detection.

Main Results:

• TRD-Net achieved an mAP@0.5 of 0.9581, a 4.32% improvement.
• Model size decreased by 19.69%, with inference time at 8.7 ms per image.
• Parameters and FLOPs reduced by 19.69% and 22.03%, respectively.

Conclusions:

• TRD-Net demonstrates significant improvements in tomato recognition and ripeness detection accuracy and efficiency.
• The model is well-suited for real-time applications in complex gardening environments.
• This research supports the development of machine vision systems for selective harvesting robots.