Fringe Projection Profilometry (FPP) is a well-known approach to estimate precise 3D profiles of objects. However, the inherent noise in fringe images hinders the reconstruction process and requires robust denoising to estimate accurate surface profiles. In this work, we proposed a Poissonian-Gaussian model for camera sensor noise in contrast to conventional Gaussian, as it is observed to be closer to actual noise distribution. Further, we developed a customized lightweight encoder-decoder network (LUNet++) with just 20K parameters to perform fringe denoising. The quantitative results on synthetic samples exhibit the proposed model’s notable 51 \(\%\) and 11 \(\%\) improvement in Peak-Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Metric (SSIM) compared to state-of-the-art methods. In addition, real results demonstrate that LUNet++ model has generalized well for real fringes, even with synthetic sample training. This effective generalization is attributed to the Poissonian-Gaussian noise model and the learning ability of the proposed LUNet++, which potentially contributes to precise surface reconstruction.