Single-shot fringe projection profilometry based on multi-scale fusion dual attention and global sliding standard deviation

Structured light fringe projection profilometry (FPP) has been widely researched and applied in industrial manufacturing, medical imaging, virtual reality, etc. However, the acquisition of high-quality fringe datasets and the depth prediction of abrupt depth regions at the edges of the single-shot fringe image are still tricky challenges. Therefore, we build a virtual FPP system to generate the desired datasets conveniently and simply based on accurately pre-calibrating intrinsics and extrinsics of the actual structured light system and propose a Multiscale Fusion Dual Attention Mechanism Network (MFDA-Unet) for the depth prediction and 3D reconstruction of a single high-frequency fringe image, combining edge gradient enhancement with global sliding standard deviation loss function to capture the detailed information in the edge regions of the objects. The training model of the network is fed back to the actual structured light system for testing, completing high-precision 3D reconstruction for complicated objects with complex surface textures from a single fringe pattern. The mean absolute error (MAE) of our proposed MFDA-UNet network is only 0.33% for the range of object depth and the whole root mean square error (RMSE) of prediction is reduced by 28.88%.