UNSUPERVISED COLOR ALIGNMENT FOR CROSS - CAMERA IMAGING APPLICATIONS

Electronics and Communication Engineering

Consistent color representation across imaging devices is crucial for numerous scientific applications. However, the inherent variations in image formation processes and the restricted access to native sensor data, particularly in consumer-grade cameras like smart phones, hinder reliable color assessment and impair computer vision performance. To address this challenge, we propose a color alignment framework that treats camera image formation as a black-box system and decomposes the color alignment task into three distinct stages: camera response calibration, response linearization and color matching. Our model can operate with non-standard color references, i.e., color patches with unknown true values, by leveraging a novel "balance-of-linear distances" feature. This approach is equivalent to determining camera parameters through an unsupervised learning process. Moreover, our method requires only a minimal set of corresponding color patches across the images to be aligned, thereby facilitating practical implementation. We evaluated our model's performance on three challenging image datasets , acquired under diverse illumination and exposure conditions by multiple cameras, including one dataset emulating uncommon scenarios encountered in scientific imaging applications. Comprehensive benchmarking demonstrated that our proposed color alignment model outperformed other popular and state-of-the-art methods, underscoring its efficacy and potential for widespread adoption.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.