Low-light RGBW Imaging demosaicking method based on residual interpolation prior and a dual-branch decoding network

With the increasing demand for low-light color imaging technology in the fields of night vision photography, military surveillance, and assistive driving, the traditional Bayer array suffers from a narrow spectral range and insufficient light intake. To overcome these limitations, the RGBW array extends the Bayer array by capturing a broader spectral range, yet its unique structure introduces a complex demosaicking challenge. In this work, a demosaicking method based on residual interpolation prior and a dual-branch decoding network is proposed for RGBW imaging. A preprocessing interpolation algorithm transforms the demosaicking task into an image restoration problem better suited for deep learning networks. The dual-branch decoding network leverages the high sensitivity of the W channel to optimize the image reconstruction. Additionally, a low-light image acquisition system is developed, and a dataset is constructed from real-world low-light scenarios. Experimental results demonstrate that the proposed method significantly improves the RGBW demosaicking performance under low-light conditions, achieving reconstructed images with enhanced details and color fidelity closely matching the human visual perception.