KAN_ICB: A Novel Deep Learning Framework for Detection of Potato Leaf Disease

Abstract

Accurate identification of potato leaf diseases is essential for timely intervention and yield protection. We present KAN ICB, a lightweight convolutional framework that embeds Kolmogorov Arnold Network basis functions with in Interactive Convolution Blocks, replacing fixed activations to capture richer nonlinear structure while preserving efficient spatial feature extraction. Using a seven class potato leaf dataset of 3,076 images with segmentation guided refinement, denoising, sharpening, standard augmentation, and resizing to 512 by 512 pixels, KAN ICB attains 99.6 percent accuracy, macro F1 of 0.98, and AUC of 0.98 on a held out test set. Model selection and variance estimation use stratified five fold cross validation on the training split, with augmentation applied only to training folds to avoid leakage. The model has 8.5 million parameters and trains in 2.5 hours on a single workstation, indicating a favorable balance between accuracy and computational efficiency. Grad CAM visualizations confirm that the network attends to lesion regions rather than background. These results show that coupling adaptive basis activations with interactive convolutions delivers high performance and deployment ready complexity for practical agricultural disease screening.