Abstract
INTRODUCTION: Cotton is a vital global economic crop and textile material, yet its yield and quality are threatened by leaf diseases such as brown spot, verticillium wilt, wheel spot, and fusarium wilt. METHODS: We propose ViTKAB, a cotton disease recognition model based on an enhanced Vision Transformer that integrates a Kolmogorov-Arnold network and a BiFormer module. The model optimizes the Vision Transformer architecture to improve inference speed, employs nonlinear feature representation to better capture complex disease characteristics, and incorporates sparse dynamic attention to enhance robustness and accuracy. RESULTS: Experiments show that ViTKAB achieves an average recognition accuracy of 98.05% across four cotton leaf diseases, outperforming models such as CoAtNet-7, CLIP, and PaLI. CONCLUSIONS: This method offers valuable insights for advancing intelligent crop disease detection systems and exhibits strong potential for deployment on edge devices.