Deep learning-based restoration method for missing fringe information in the high reflectivity regions

Abstract

Conventional fringe projection profilometry (FPP) based on a single-exposure is limited in achieving high-precision 3D measurements when processing fringe images that contain saturated regions, as distortion in these areas significantly reduces the measurement accuracy. Although high dynamic range (HDR) methods can mitigate this issue, they require additional conditions, which increase both measurement costs and complexity. To address the limitations of traditional HDR methods, a method based on deep learning is proposed for restoring missing fringe information. The proposed method employs a simple U-Net architecture to efficiently restore saturated fringe images by leveraging high-dimensional feature representations and skip connections within the network. In addition, combining the restored fringe information provided by the method presented enables 3D measurement of the object under different measurement systems. The results of the experiments confirm that the method accurately restores the fringe information in saturated images, making the restored images suitable for high-precision 3D reconstruction. Furthermore, by integrating the restored fringe information, the proposed method enables 3D reconstruction using saturated fringe images captured under different measurement systems, which include those with varying saturation levels and those captured from different angles. This demonstrates that the proposed method restores missing fringe information in saturated images and exhibits strong generalization capabilities.