With the rapid development of high-speed image sensors and optical imaging technology, these have effectively promoted the improvement of non-contact 3D shape measurement. Among them, striped structured-light technology has been widely used because of its high measurement accuracy. Compared with classical methods such as Fourier transform profilometry, many deep neural networks are utilized to restore 3D shape from single-shot structured light. In actual engineering deployments, the number of learnable parameters of convolution neural network (CNN) is huge, especially for high-resolution structured-light patterns. To this end, we proposed a dual-path hybrid network based on UNet, which eliminates the deepest convolution layers to reduce the number of learnable parameters, and a swin transformer path is additionally built on the decoder to improve the global perception of this network. The experimental results show that the learnable parameters of the model are reduced by 60% compared with the UNet, and the measurement accuracy is not degraded at the same time. The proposed dual-path hybrid network provides an effective solution for structured-light 3D reconstruction and its practice in engineering.

Structured light profilometry (SLP) is now widely utilized in noncontact threedimensional (3D) reconstruction due to its convenience in dynamic measurements. Compared with classic fringe projection profilometries, multiple deep neural networks are proposed to demodulate or unwrap the fringe phase, and these networks utilize convolution layers to extract local features while omitting global characteristics. In this paper, we propose SwinConvUNet, a deep neural network for single-shot SLP that can extract local and global features simultaneously. In the network structure design, convolution layers are applied in shallow layers to extract local features, whereas transformer layers extract global features in deep layers, and an improved loss function by combining gradient-based structural similarity is employed to improve reconstruction details. The experimental results demonstrate that SwinConvUNet is more effective than the U-net model at decreasing learnable parameters while maintaining 3D reconstruction accuracy.

With the rapid development of high-speed image sensors and optical imaging technology, these have effectively promoted the improvement of non-contact 3D shape measurement. Among them, striped structured-light technology has been widely used because of its high measurement accuracy. Compared with classical methods such as Fourier transform profilometry, many deep neural networks are utilized to restore 3D shape from single shot structured-light. In actual engineering deployments, the number of learnable parameters of convolution neural network (CNN) is huge, especially for high-resolution structured-light patterns. To this end, we proposed a dual-path hybrid network based on UNet, which eliminates the deepest convolution layers to reduce the number of learnable parameters, and a swin transformer path is additionally built on the decoder to improve the global perception of this network. The experimental results show that the learnable parameters of the model are reduced by 60% compared with the UNet, and the measurement accuracy is not degraded at the same time. The proposed dual-path hybrid network provides an effective solution for structured-light 3D reconstruction and its practice in engineering.