HNet: Multi-Modal Dynamic Fusion Model for Detecting Hidden Crops in Fields

In the domain of agricultural intelligence, covert crop detection is essential for safeguarding crop growth and enhancing the efficiency of agricultural production. Traditional camouflaged object detection (COD) methodologies are primarily designed for large-scale targets such as animals and humans. However, these methods often demonstrate inadequate featurecapture capabilities and are prone to background interference when applied to small and heavily occluded crops in agricultural environments. Although the Recurrent Iterative Segmentation Network (RISNet) [1] enhances dense target detection through hierarchical progressive optimization, the feature discrimination capacity of the existing methods remains insufficient for smallsized crops in complex occlusion scenarios, and the model optimization process is characterized by redundancy. This study introduces the HNet model, which facilitates crossscale feature interaction through a cross-modal dynamic feature fusion module (CMF) and enhances the extraction of irregular leaf features by integrating it with a direction-aware dynamic gating module (DGM). Experimental results show that (HNet) maintains an S-measure of 86.7 % in heavy occlusion scenarios and achieves stable convergence within a limited training period, providing a more reliable solution for intelligent monitoring of complex farmland environments.