Learned NIR&VISCAM: multi-spectral fusion for large depth-of-field computational imaging

Abstract

Large DOF (depth-of-field) with high SNR (signal-noise-ratio) imaging plays an important role in many applications such as unmanned driving to medical imaging. However, there is always a trade-off between DOF and SNR in traditional optical design. In this paper, we propose a NIR&VISCAM (NIR&VIS Camera) that combines multi-spectral optical design and deep learning to realize large DOF and high SNR imaging. Specifically, a multi-spectral optical imaging system based on the HVS (human visual system) is designed to provide colorful but small DOF VIS (visible) image and large DOF NIR (near-infrared) image. To achieve DOF extension, we build a fusion network NIR&VISNet consisting of a VIS encoder for color extraction, a NIR encoder for spatial details extraction and a decoder for information fusion. We establish a prototype to capture real-scene dataset containing 1000 sets and test our method on a variety of test samples. The experimental results demonstrate that our NIR&VISCAM can effectively produce large DOF images with high quality. Moreover, compared to the classic image fusion methods, our designed algorithm achieves the optimal performance in DOF extension and color fidelity. With the prominent performance in large DOF and high SNR imaging, this novel and portable system is promising for vision applications such as smartphone photography, industry detection, and life medical.