Efficient Quad Bayer Demosaicing with Look-Up Tables

ICLR 2026 Conference Submission17972 Authors

Keywords: quad bayer, demosaicing, real-time, efficient, isp

Abstract: This paper presents a novel look-up table-based demosaicing method for quad Bayer pattern. Recent deep learning approaches, though effective, are computationally expensive and unsuitable for efficient implementation. We first introduces a residual LUT to reduce memory and computational complexity. The method employs a two-stage demosaicing process: the first stage performs primary demosaicing, while the second stage enhances high-frequency components. Each stage consists of a number of residual LUTs and they are stacked both in serial and parallel to effectively enlarge the receptive field size. In experiments, our method achieves stable performance and improved image quality with efficiency. In addition, the size and the computational complexity of our method enable efficient hardware implementation.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 17972