RU 
EN 
CN 
ES 
Scientists from the Institute of Artificial Intelligence at Samara University have developed a new technology for improving the color accuracy of digital cameras. The work was carried out jointly with colleagues from Moscow, Würzburg (Germany) and York (Canada). For the first time ever, scientists used Kolmogorov-Arnold networks (KANs) to automatically process camera images. This is a new class of neural networks that allowed neural network methods to outperform classical color processing methods.
The developed technology was tested on several classes of tasks related to color image processing. Test results have shown that the solution called cmKAN significantly outperforms various color enhancement methods currently used by smartphone and digital camera manufacturers worldwide.
“In the modern world, people use a wide variety of cameras; modern smartphones, as a rule, operate three types of cameras – the TV camera, the main camera, and the wide-angle camera. Different cameras display color in dissimilar way; when switching between cameras, the colors in the images vary significantly; moreover, these differences between cameras are nonlinear, which complicates both perception and automatic image processing. Despite the impressive successes of neural network methods in virtually all areas of computer vision, accurate color processing has remained the domain of classical algorithms thus far. This is apparently due to the high sensitivity of human color perception, as well as the fact that modern neural network approaches do not adequately capture the specifics of color transformation. We've bridged the gap between classical and neural network color processing algorithms and developed a universal neural network color matching approach, cmKAN, which enables more accurate comparison and automatic color correction of images. The neural network's operating process is similar to the steps taken by an operator of a color correction program such as Photoshop or Lightroom. The operator constructs nonlinear color transformation curves and specifies their coverage areas: for example, different color transformation rules apply to the bright sky, the shadows of buildings, and the vicinity of light sources. Our approach works in a similar way. We were able to theoretically demonstrate that Kolmogorov-Arnold networks best reflect nonlinear color transformations, and the parameters of these transformations in different parts of the image are determined by a generator network,” says Professor Artem Nikonorov, Director of the Institute of Artificial Intelligence and Head of the Centre for Intelligent Mobility of Multifunctional Unmanned Aerial Systems at Samara University.
Kolmogorov-Arnold networks (KANs) are a new type of neural network architecture based on the Kolmogorov-Arnold representation theorem developed by Soviet mathematicians Andrey Kolmogorov and Vladimir Arnold. This type of neural network architecture was developed in 2024 and could become an alternative to traditional MLP (multilayer perceptron) neural networks, which are now widely used in computer vision systems and large language models.
To train and test cmKAN, the developers prepared and published an impressive Volga2K dataset containing over two thousand pairs of images from various cameras, in various locations and shooting conditions.
“Our technology was tested for the main tasks of color image transformation: color synchronization of images from two different cameras; color synchronization of RAW images from different cameras; conversion of RAW images to the final image, which is the main task for any modern camera, both on smartphones and professional ones; and, most importantly, human post-processing of images. The results showed that our method consistently outperforms global analogues by an average of 37.3 %. In addition, cmKAN effectively handles high dynamic range scenes, images captured in challenging conditions, and images taken in the evening and at night. Our method can be used not only in smartphone cameras, but also in the creation of new image processors for digital cameras, as well as in the automation of photo editing and color correction in publishing, printing, media content preparation, and professional visualization,” notes Artem Nikonorov.
The developed cmKAN approach was presented at the IEEE International Conference on Computer Vision, which was held from October 19 to 23, 2025, in Honolulu, Hawaii.
The project was developed by scientists from Samara National Research University, Moscow Institute of Physics and Technology, Research Institute of Artificial Intelligence (Moscow), A.A. Kharkevich Institute for Information Transmission Problems of the Russian Academy of Sciences (Moscow), University of Würzburg (Germany), and York University (Canada). The project was implemented as part of the work of the Artificial Intelligence Research Centre – the Centre for Intelligent Mobility of Multifunctional Unmanned Aircraft Systems.
The material is prepared with the support of Russia’s Ministry of Education and Science, in the framework of the Decade of Science and Technology.
