Y-FFC Net for 3D Reconstruction of Highly Reflective Surfaces
Xuwen Song, Lianpo Wang
Abstract
3D reconstruction plays a pivotal role in intelligent manufacturing, industrial inspection, and other fields. Fringe projection profilometry is a widely used 3D reconstruction method due to its high accuracy and noncontact nature. However, when fringe projection profilometry (FPP) is applied to reconstruct highly reflective surfaces, there are not only diffuse reflections, but also undesired specular reflections. Intense specular reflection destroys the sinusoidal characteristics of the fringe pattern, leading to a decrease in 3D reconstruction accuracy. Traditional methods, such as Multiple exposures method, require multiple redundant shots to obtain high-quality fringe patterns, which makes efficiency difficult to meet industrial measurement requirements. Inspired by the application of deep learning in 3D vision, this article proposes a Y-shaped fast Fourier convolutional network (Y-FFC) for high reflection removal to provide high-quality fringe patterns. The cosine characteristics of the grayscale gradient of the sinusoidal fringe pattern make the detection of nonsinusoidal regions more accurate in the gradient domain. In addition, the undesired specular reflection component is easier to distinguish and filter out in the frequency domain. Therefore, the design of Y-FFC fully considers the gradient and frequency information. Experimental results show that the introduction of gradient and frequency information is beneficial for high reflection removal, enabling the reconstruction of industrial workpieces with highly reflective surfaces without the need for additional shots. The mean absolute error for depth measurements of an aircraft blade with a depth of 30 mm was reduced from 0.1332 to 0.0041 mm.