Matching color characteristics of images using Kolmogorov-Arnold neural networks based on hypernetwork architecture

cmKAN, a universal color matching framework, has been introduced. The proposed method makes it possible to efficiently and accurately display the colors of the input image from the original color distribution to the target one in both controlled and uncontrolled settings.
The framework is based on the use of spline capabilities of Kolmogorov-Arnold (KAN) networks to model color matching between source and target distributions. In particular, a hypernet has been developed that generates spatially variable weight maps to control the nonlinear splines of the KAN network, which ensures accurate color matching. This paper also presents a large-scale dataset consisting of paired images obtained from two different cameras and designed to evaluate the effectiveness of the method in the task of matching colors created by different cameras.
The approach was evaluated on various color matching tasks, including: (1) mapping from RAW to RAW, where the original color distribution is in the RAW space of one camera and the target color distribution is in the RAW space of another; (2) mapping from RAW to sRGB, where the original distribution is in the RAW space of the camera, and the target is in the sRGB space for displays, which emulates the color rendering of the camera's image processing processor (ISP); and (3) sRGB-to-sRGB mapping, which aims to transfer colors from the sRGB source space (for example, obtained with the ISP of one camera) to the sRGB target space (for example, from the ISP of another camera).
The results obtained demonstrate that the proposed method achieves state-of-the-art performance in all the tasks considered, while remaining computationally lightweight compared to other methods of color matching and transfer.