Robust and real-time 3D reconstruction with quaternary & octonary gray code fringe projection profilometry and AI-based noise suppression

Abstract

Fringe Projection Profilometry (FPP) was an established technique for non-contact 3D surface reconstruction. This research investigated advancements in FPP, focusing on enhancing both acquisition speed and reconstruction quality. Specifically, the study explored the implementation of higher-base Gray Code (GC) unwrapping methods and the application of Artificial intelligence (AI) for image enhancement.

Octonary Gray Code (OGC) and the Nero AI filter were combined to successfully achieve OGC unwrapping, a task previously considered infeasible. To accelerate acquisition, Quaternary Gray Code (QGC) and Octonary Gray Code (OGC) unwrapping techniques were evaluated, aiming to reduce the number of projected patterns. Real-time object capture was conducted using QGC, and processing efficiency was optimized through algorithmic refinement and Graphical Processing Unit (GPU) acceleration. This optimization resulted in a significant reduction in processing time, achieving an 87.6% decrease from 13.52 seconds to 1.68 seconds for 100 cycles.

Furthermore, the study examined the application of AI-driven image enhancement to improve reconstruction quality. Several AI models, including the Residual Dense Swin Transformer Network (RDSTN), OpenVINO Enhanced Deep Super-Resolution Network (EDSR), Nero AI Image Denoiser, and Canva AI Photo Enhancer, were compared for noise suppression and resolution enhancement. Results demonstrated substantial improvements in both metrics. Notably, the Nero AI Image Denoiser exhibited superior performance, achieving a 93.7% reduction in unwrapping noise, from a maximum absolute error of 200.9 radians to 12.6 radians. This method also preserved finer object details compared to conventional median filtering, which often compromised edge and depth information during noise reduction. In conclusion, this research demonstrated the efficacy of employing higher-base GC and AI-based image processing to significantly enhance the speed and accuracy of FPP, presenting a more robust and efficient 3D reconstruction methodology.