Color Image Demosaicking: A Systematic Survey of Algorithms, Performance Evaluation, and Open Challenges

The color filter array (CFA) is a crucial component in digital cameras, which segregates various color wavelengths by overlaying the sensor surface with specific color filters. This design enables the collection of color information from the captured scenes. Among the different types of CFAs, Bayer CFA is the most prevalent. Demosaicking refers to the process by which missing color information is estimated using algorithmic techniques, leveraging existing color data to reconstruct full-resolution images from samples obtained via CFA sensors. Traditional demosaicking approaches typically fall into three categories: interpolation-based, Bayesian-based and frequency domain-based algorithms. In contrast, learning-based methods encompass both dictionary learning and deep-learning frameworks. This survey presents a comparative analysis of representative techniques, focusing on their efficacy in terms of texture and color restoration accuracy evaluated against the Kodak, McMaster, and DIV2K-RGB subset datasets. The experiment results show that the demosaicking effect of color images based on learning-based methods is generally better than that of traditional methods. For instance, compared with the representative ARI in traditional methods, in the three datasets, the average PSNR of the best deep learning method, RSTCANet, is 2.59db higher, the SSIM is 0.0081 higher, and the CIEDE2000 is 0.1840 lower. Furthermore, we address the challenges associated with contemporary demosaicking methodologies and explore the future trends in this area. This survey serves as a valuable resource for novices interested in color image demosaicking algorithms and researchers experienced in this domain.