Dynamic Coding Network for Robust Fruit Detection in Low-Visibility Agricultural Scenes

Hanyun Lu¹Teng Jin¹Chen Wan¹Shuifa Sun²Xiumei Zhou^3*Fangyi Wang^1*

• ¹College of Computer and Information Technology, China Three Gorges University, Yichang 443002, China, yichangshi, China
• ²School of Information Science and Technology, Hangzhou Normal University, Hangzhou, 311121, Zhejiang, China, Hangzhou, China
• ³Innovation and Entrepreneurship College of China Three Gorges University, Hubei, Yichang, 443002, China, yichang, China

Accurate fruit detection under low-visibility conditions is essential for intelligent orchard management and robotic harvesting. We propose a dynamic coding-based detection network that integrates a dynamic feature encoder , a global attention gate , an lterative feature attention module, and a cross-attention decoder to address feature degradation, background clutter, and fine-grained detail preservation in challenging agricultural environments. Experiments on the LVScene4K dataset, which includes multiple fruit categories (grape, kiwifruit, orange, pear, pomelo, persimmon, pumpkin, and tomato) captured under fog, rain, low light, and occlusion conditions, demonstrate that DCNet achieves 86.5% mean average precision and 84.2% intersection over union . Compared with state-of-the-art baselines, DCNet improves F1 by 3.4% and IoU by 4.3% while maintaining an inference speed of 28 FPS on an RTX 3090 GPU. These results confirm that DCNet provides a superior balance between accuracy and efficiency, making it well-suited for real-time deployment in agricultural robotics.Code is available at: https://github.com/PAIGUI/DCNet.git.