Abstract
High-quality disease segmentation plays a crucial role in the precise identification of rice diseases. Although the existing deep learning methods can identify the disease on rice leaves to a certain extent, these methods often face challenges in dealing with multi-scale disease spots and irregularly growing disease spots. In order to solve the challenges of rice leaf disease segmentation, we propose KBNet, a novel multi-modal framework integrating language and visual features for rice disease segmentation, leveraging the complementary strengths of CNN and Transformer architectures. Firstly, we propose the Kalman Filter Enhanced Kolmogorov-Arnold Networks (KF-KAN) module, which combines the modeling ability of KANs for nonlinear features and the dynamic update mechanism of the Kalman filter to achieve accurate extraction and fusion of multi-scale lesion information. Secondly, we introduce the Boundary-Constrained Physical-Information Neural Network (BC-PINN) module, which embeds the physical priors, such as the growth law of the lesion, into the loss function to strengthen the modeling of irregular lesions. At the same time, through the boundary punishment mechanism, the accuracy of edge segmentation is further improved and the overall segmentation effect is optimized. The experimental results show that the KBNet framework demonstrates solid performance in handling complex and diverse rice disease segmentation tasks and provides key technical support for disease identification, prevention, and control in intelligent agriculture. This method has good popularization value and broad application potential in agricultural intelligent monitoring and management.