Three-dimensional imaging plays a crucial role at the micro-scale in fields such as precision manufacturing and materials science. However, image noise significantly impacts the accuracy of point cloud reconstruction, making image denoising techniques a widely discussed topic. Statistical analysis of laser imaging noise has led to the conclusion that logarithmically transformed noise follows a Gumbel distribution. A corresponding neural network training set was developed to address the challenges of difficult data collection and the scarcity of phase-unwrapped image datasets. Building on this foundation, a phase-unwrapped image denoising method based on the Denoising Convolutional Neural Network (DnCNN) is proposed. This method aims to achieve three-dimensional filtering by performing two-dimensional image denoising. Experimental results show a significant reduction in the Cloud-to-Mesh Distance (C2M) statistics of the corresponding point clouds before and after planar filtering. Specifically, the statistic at 97.5% of the 2σ principle decreases from 0.8782 mm to 0.3384 mm, highlighting the effectiveness of the filtering algorithm in improving the planar fit. Moreover, the DnCNN method exhibits exceptional denoising performance when applied to real-world target data, such as plaster statues with complex depth variations and PCBs made from different materials, thereby enhancing accuracy and reliability in point cloud reconstruction. This study provides valuable insights into phase-unwrapped image noise suppression in laser imaging, particularly in micro-scale applications where precision is critical.
Read full abstract