Inferring Objects From Lensless Imaging Measurements

Xiangjun Yin, Huanjing Yue, Mengxi Zhang, Huihui Yue, Xingyu Cui, Jing-Yu Yang

Abstract: Various camera-based object inference applications such as Internet of Things (IoT), have critical restrictions on the size and weight of the cameras. Emerging lensless cameras offer an increasingly promising option in terms of lightness. This paper presents the first attempt towards pixel-wise lensless object inference (LOI). Specifically, we propose an end-to-end LOI network (LOINet) consisting of three phases: (1) To acquire scene-resembling feature representations, we design the spatial feature learning (SFL) phase; (2) To highlight object regions from features for inference, we construct the positioning phase by a contextual exploration module (CEM) and a hierarchical feature fusion module (HFFM); (3) To remove false-positive/negative features for accurate inference, we propose the focusing phase consisting of a series of hierarchical cascading dual-branch attention (DBA). Furthermore, we construct the first dataset for investigation of LOI. Extensive experiments demonstrate that our method achieves high accuracy, which sheds light on high-level inference from lensless measurements.