SMART-NIR: A Multi-Kernel Vision Transformer with Kolmogorov-Arnold Network Classifier for Near-Infrared Spectral Classification

Near-infrared (NIR) spectroscopy is a powerful analytical tool widely applied in classification and quality control. However, it faces significant challenges when handling complex, multidimensional, and nonlinear data. This paper introduces SMART-NIR, an innovative framework designed to enhance NIR spectral analysis by integrating multi-kernel feature extraction, an improved Vision Transformer architecture with Dual-MLP, and Kolmogorov-Arnold Networks (KAN) for superior classification performance. The proposed framework achieved an impressive classification accuracy of 99.24%, surpassing traditional Transformer models. The multi-kernel feature extraction technique allows adaptive extraction of key spectral features, while KAN integration improves classification accuracy by 5% over standard MLP classifiers. Additionally, the Dual-MLP architecture significantly reduces the number of parameters and floating point operations (FLOPs) compared to conventional feed-forward networks used in Transformers. SMART-NIR demonstrates exceptional capability in capturing complex, nonlinear data relationships, making it highly effective for real-time inspection and analysis in dynamic, noisy environments. The framework offers a promising solution for enhancing the accuracy and efficiency of NIR spectroscopy across various practical applications.