An effective citrus ripeness detection model for complex orchard scenarios

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Detecting citrus ripeness in orchards is challenging due to obstructions, uneven lighting, and overlapping fruits.
• Existing models struggle with accuracy and efficiency in complex orchard environments.

Purpose of the Study:

• To develop an effective and efficient citrus ripeness detection model for orchard environments.
• To address challenges like leaf obstructions, uneven illumination, and small target detection.

Main Methods:

• Proposed the DC-YOLO model integrating dynamic depth-separable convolutions and a collaborative attention mechanism.
• Utilized dynamic depth-separable convolutions to reduce complexity and enhance feature extraction.
• Implemented a collaborative attention module (coordinate, cross-scale, dual-feature multi-head self-attention) for improved feature fusion and localization.
• Adopted a multi-loss function for localization, classification, ripeness assessment, and occlusion awareness.

Main Results:

• Achieved high performance with precision (0.980), recall (0.965), F1-score (0.972), mAP50 (0.975), and mAP50:95 (0.769).
• Reduced model parameters by 12.79% and computational complexity by 18.28% compared to baseline models.
• Reached an inference speed of 93.42 FPS, balancing accuracy and real-time performance.

Conclusions:

• The DC-YOLO model demonstrates strong potential for real-time citrus ripeness detection in orchards.
• Further research is needed for edge device deployment and handling diverse citrus data samples.