Dynamic coding network for robust fruit detection in low-visibility agricultural scenes.

Area of Science:

• Computer Vision
• Agricultural Technology
• Robotics

Background:

• Accurate fruit detection is vital for automated agriculture, but low-visibility conditions (fog, rain, low light) degrade existing model performance.
• Intelligent orchard management and robotic harvesting require robust fruit detection systems that can overcome environmental challenges.

Purpose of the Study:

• To develop a novel, modular detection framework, the Dynamic Coding Network (DCNet), for enhanced fruit detection in low-visibility agricultural scenes.
• To address the performance degradation of current models in visually challenging environments.

Main Methods:

• Proposed DCNet framework with four key components: Dynamic Feature Encoder, Global Attention Gate, Cross-Attention Decoder, and Iterative Feature Attention.
• Utilized the LVScene4K dataset featuring diverse fruits (grape, kiwifruit, orange, etc.) under various adverse conditions (fog, rain, low light, occlusion).

Main Results:

• DCNet achieved 86.5% mean average precision (mAP) and 84.2% intersection over union (IoU) on the LVScene4K dataset.
• Outperformed state-of-the-art methods with a 3.4% F1 score and 4.3% IoU improvement.
• Maintained real-time inference speed of 28 FPS on an RTX 3090 GPU.

Conclusions:

• DCNet offers a superior balance of accuracy and computational efficiency for real-time agricultural robotics deployment.
• The modular architecture of DCNet demonstrates potential for generalization across different crops and complex agricultural settings.