Abstract
In precision agriculture, effective weed management is pivotal for enhancing rice cultivation yield and quality. However, accurately differentiating weeds from rice crops remains a fundamental challenge for precision weeding. This study introduces an innovative deep-learning methodology based on style transfer for weed identification and segmentation in paddy fields. We introduce a Style-guided Weed Instance Segmentation (SWIS) method that integrates a Random Adaptive Instance Normalization (RAIN) module for stochastic style transformation and a Dynamic Gradient Back-propagation (DGB) module for adversarial feature optimization. Specifically, the RAIN module aligns feature distributions between laboratory and field environments through stochastic style transformation, enhancing cross-environment generalization. The DGB module employs adversarial optimization with gradient-guided perturbations to enhance feature robustness under complex field conditions. Experimental results demonstrate that our method achieves a Weed Intersection over Union (Weed IoU) of 70.49% on field data, significantly outperforming comparison methods. Therefore, this approach proves effective for real-world applications. Beyond its immediate applications, this research advances computer vision integration in agriculture and establishes a robust foundation for developing more sophisticated, versatile weed recognition models.