Photon-Efficient 3-D Imaging via a Lightweight Pyramid Network

Photon-efficient imaging founded on single-photon light detection and ranging (LiDAR) has significantly broadened the scope of human vision. Nevertheless, two major challenges impede its potential lightweight applications: the design of an efficient and effective architecture and the correlation extraction mechanism. In this article, we pioneer a novel lightweight pyramid network that, for the first time, tackles both challenges concurrently, thereby enabling lightweight photon-efficient imaging. First, we delve into the disparities in the histogram distributions of photon-counting cubes at different frequencies and utilize pyramid decomposition to streamline the learning process. Simultaneously, we analyze the prominent signal clustering characteristics in cross-dimensional views and adapt the triplet attention module (TAM) to extract spatio-temporal correlations. Evaluations on simulated data demonstrate that, despite being 76% smaller in model size compared to the smallest existing network, the proposed network attains state-of-the-art photon-efficient imaging performance. Moreover, our network exhibits satisfactory generalization capabilities for real-world scenes captured by indoor and outdoor imaging systems. This accomplishment paves the way for the further expansion of photon-efficient imaging in real-time applications.