MFENet: Multi-Frequency Edge Detection Network for Plant Camouflage

本文针对植物伪装检测任务中低信噪比、动态干扰与细粒度识别问题，提出一种多频边缘动态检测网络MFENet，以提升复杂生态场景下的隐蔽植物检测精度与鲁棒性。在MFENet中，采用多尺度频率分离模块处理低信噪比问题，通过多尺度分组卷积分离低频全局特征与高频细节特征。同时，构建了通道感知边缘注意力模块，结合Sobel边缘先验与通道–空间注意力优化细粒度特征。为进一步提升检测精度，提出了边缘强度驱动的动态迭代反馈机制，自适应调整计算复杂度。在PlantCamo数据集上，与通用的伪装检测模型相比，MFENet在各指标上提升明显。消融实验验证了各模块的有效性。MFENet显著提升了植物伪装检测的精度与效率，为生态保护与农业监测提供可靠技术支撑。This paper addresses the problems of low signal-to-noise ratio, dynamic interference, and fine-grained recognition in the task of plant camouflage detection. We propose a multi-frequency edge dynamic detection network (MFENet) to enhance the accuracy and robustness of concealed plant detection in complex ecological scenes. MFENet employs a multi-scale frequency separation module to address the low signal-to-noise ratio issue by utilizing multi-scale grouped convolutions to separate low-frequency global features from high-frequency detail features. Additionally, we construct a Channel-Aware Edge Attention module that combines Sobel edge priors with channel-space attention to optimize fine-grained features. To further enhance detection accuracy, we introduce an edge intensity-driven dynamic iterative feedback mechanism to adaptively adjust computational complexity. On the PlantCamo dataset, MFENet shows significant improvements in all evaluation metrics compared to conventional camouflage detection models. Ablation studies validate the effectiveness of each module. MFENet significantly enhances the accuracy and efficiency of plant camouflage detection, providing reliable technical support for ecological protection and agricultural monitoring.