LVR: A language and vision fusion method for rice diseases segmentation under complex environment

Abstract

Accurate identification of rice diseases depends on high-quality disease segmentation. However, challenges such as the complexity of the rice field environment, interference from redundant information, and slow model convergence can hinder effective segmentation. To address these issues, we propose A Language and Vision Fusion Method for Rice Diseases Segmentation under complex environment (LVR), which combines CNN and Transformer architectures. First, we present the Efficient Wavelet-based Multi-scale Attention (EWWL) module, designed to enhance the model’s ability to capture fine details of disease regions in complex environments. Next, to mitigate information redundancy, we design the KAN-segmentation (KAN-seg) module for efficient feature extraction. Additionally, we propose a Self-Adaptive Gradient Enhancement (SAGE) algorithm that dynamically adjusts the network’s learning rate, thereby accelerating convergence. Experimental results demonstrate that the LVR method achieves exceptional accuracy and robustness in rice disease segmentation, even under challenging field conditions. This provides substantial technical support for intelligent agricultural disease management and offers promising applications, particularly in the realm of smart agricultural disease monitoring and management.