Abstract: With the advancement of deep learning, single-photon imaging has gradually become an important and challenging research direction. It contributes to the 3D reconstruction of single-photon images, which are sparse, noise-filled 3D images with only a few valid signal echoes in their time channels. However, current architectures generally achieve better results by establishing larger backbone networks, which comes at the cost of higher GPU memory usage. Designing a lightweight yet effective model that can be deployed on edge devices is also a significant concern in the research community. Therefore, this paper proposes a lightweight architecture that achieves comparable results to other methods with significantly reduced computational requirements. Specifically, this work utilizes a Swin Transformer network to extract temporal features of single-photon images, using this time prediction network to significantly reduce the dimensions of single-photon images. Then, a Densely Cascaded Multi-scale Network (DCMNet) is employed to extract spatial domain features of single-photon images for further reconstruction. It improves the interconnection between decoding layers through a top-down cascade pathway and dense connections, producing high-quality multi-resolution depth outputs. Experimental results demonstrate that our network can achieve commendable results while significantly reducing resource consumption.

Key words: computer vision, three-dimensional reconstruction of single-photon images, Swin Transformer and dense cascaded multi-scale network, lightweight architecture, edge computing