Three-dimensional Shape Reconstruction from Single-shot Speckle Image Using Deep Convolutional Neural Networks

Three-dimensional (3D) shape reconstruction from a monocular two-dimensional (2D) image has emerged as a highly demanded tool in many applications. This paper presents a novel 3D shape reconstruction technique that employs an end-to-end deep convolutional neural network (CNN) to transform a single speckle-pattern image into its corresponding 3D point cloud. In the proposed approach, three CNN models are explored for comparison to find the best capable network. To train the models with reliable datasets in the learning process, a multi-frequency fringe projection profilometry technique is adopted to prepare high-accuracy ground-truth 3D labels. Unlike the conventional 3D imaging and shape reconstruction techniques which often involve complicated algorithms and intensive computation, the proposed technique is simple, yet very fast and robust. A few experiments have been conducted to assess and validate the proposed approach, and its capability provides promising solutions to the ever-increasing scientific research and engineering applications.