Kolmogorov-Arnold Networks in Transformer Attention for Low-Light Image Enhancement

Recent advancements in deep learning, particularly in Transformer architectures, have revolutionized tasks that require processing complex, highly-detailed data. One such task is low-light image enhancement, a critical sub-field of image restoration, which has benefited from extensive research into various Transformer models, each offering distinct strengths. This paper introduces a novel Transformer attention mechanism inspired by the Kolmogorov - Arnold representation theorem, leveraging learnable non-linearity and multivariate function de-composition. This attention mechanism is utilized to develop KAN - T, our Transformer network. These enhancements enable greater feature flexibility and broader context capture within the model. Our comprehensive experiments, both quantitative and qualitative, demonstrate that the proposed method achieves state-of-the-art performance in low-light image enhancement, highlighting its effectiveness and wide-ranging applicability.