Lightweight and edge-preserving speckle matching network for precise single-shot 3D shape measurement. (31st March 2023)

• Record Type: Journal Article
• Authors: Dong, Yanzhen; Yang, Xiao; Wu, Haitao; Chen, Xiaobo; Xi, Juntong
• Is Part Of: Measurement. Volume 210(2023)
• Publication Date: 2023-03-31
• Subjects: Single-shot shape measurement -- Speckle matching network -- Edge-preserving -- Digital speckle correlation; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2023.112549
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A lightweight and edge-preserving speckle matching network is proposed for precision measurement. High-precision 3D shape can be reconstructed from a single-shot speckle image. The proposed network outperforms the traditional digital speckle correlation in inference speed and data integrity. Compared with other networks, proposed speckle matching network predicts disparity maps with greater precision while occupying less computational consumption. A precise and reliable speckle binocular dataset based on DSC is built using a specialized device. Abstract: Single-shot three-dimensional (3D) shape measurement method based on digital speckle correlation (DSC) has high measurement precision and only requires one pair of speckle images. Nevertheless, the matching of speckles is time-consuming and may lead to some data defects at large gradient variations and edge regions. This paper proposes an end-to-end speckle matching network to achieve fast, precise, and edge-preserved 3D measurement. Compared with other networks, the proposed network reduces the redundant channels in the Siamese feature-extraction subnetwork and constructs cost volume by group-wise correlation, making it possible to provide efficient representations at low computational consumption, thus being suitable for high-precision 3D measurement in large resolution. A lightweight 3D stacked hourglass subnetwork is used to aggregate cost information and furthermore, an edge refinement module is integrated by leveraging the left speckle image as a guide to regular edge details. A reliable speckle binocular dataset is prepared with a specialized device in a simple procedure and the labels are produced from the DSC method. Experiments demonstrate that the proposed network can produce dense disparity maps with subpixel precision and reduce single-shot inference time to 0.13 s with less GPU occupation, which is a significant improvement compared with the traditional DSC method and other learning-based methods. Moreover, the experiment validates that measurement precision of the reconstructed plane can reach less than 0.03 mm, meeting the requirements of most industrial applications.