Brateanu Alexandru, Balmez Raul, Orhei Ciprian, Ancuti Cosmin, Ancuti Codruta
Department of Computer Science, University of Machester, Manchester M13 9PL, UK.
Faculty of Electronics, Telecommunications and Information Technologies, Polytechnic University Timisoara, 300223 Timisoara, Romania.
Sensors (Basel). 2025 Jan 8;25(2):327. doi: 10.3390/s25020327.
Low-light image enhancement (LLIE) techniques improve the performance of image sensors by enhancing visibility and details in poorly lit environments and have significantly benefited from recent research into Transformer models. This work presents a novel Transformer attention mechanism inspired by the Kolmogorov-Arnold representation theorem, incorporating learnable non-linearity and multivariate function decomposition. This innovative mechanism is the foundation of KAN-T, our proposed Transformer network. By enhancing feature flexibility and enabling the model to capture broader contextual information, KAN-T achieves superior performance. 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. The code will be released upon publication.
低光图像增强（LLIE）技术通过增强光线不足环境中的可见性和细节来提高图像传感器的性能，并且从最近对Transformer模型的研究中受益匪浅。这项工作提出了一种受柯尔莫哥洛夫-阿诺德表示定理启发的新型Transformer注意力机制，融合了可学习的非线性和多元函数分解。这种创新机制是我们提出的Transformer网络KAN-T的基础。通过增强特征灵活性并使模型能够捕获更广泛的上下文信息，KAN-T取得了卓越的性能。我们全面的定量和定性实验表明，所提出的方法在低光图像增强方面达到了当前的最优性能，突出了其有效性和广泛的适用性。代码将在发表时发布。
