A method of laser stripe centerline extraction based on deep learning for structured light 3D reconstruction
Ming Huang, Xiaohan Xu
Abstract
Abstract To overcome the stray light noise in the centerline extraction method during line structured light 3D reconstruction process, an end-to-end trainable neural network for laser stripe centerline extraction based on Convolutional Neural Network and Multi-Layer Perception is proposed. The proposed network can self-adapt to a variety of lighting (brightness) conditions and overcome the interference of different stray lights. In addition, unlike prior work on enhancing the accuracy of centerline extraction using deep learning methods that only performs it for noise reduction in pre-processing, the proposed network unifies the noise reduction and prediction processes, so that it can be optimized end-to-end directly on centerline extraction performance. The network learns an intermediate feature representation of noise reduction, which requires less complexity for data annotation, reduces the training difficulty, and has more scalability. Experiments show that the proposed method can perform centerline extraction with relatively high accuracy for laser stripes of different widths, brightness, and inclination, thus obtaining a smooth and stable reconstructed surface in the structured light 3D reconstruction process.