Shape reconstruction from focus detection and signal classification using an optical microscope

Abstract

Shape reconstruction from focus using a microscope is a cost-effective technique for restoring three-dimensional shapes, making it well-suited for high-precision measurement needs at microscopic scales. However, its effectiveness is limited by the variability of focus-measured signals and the alignment accuracy between corresponding image points. To address these limitations, a novel shape reconstruction method is proposed that integrates focus detection, the classification of focus-measured signals, and depth reconstruction. This method employs deep learning to categorize the geometric shapes of focus-measured signals, identify those characterized by noise, and perform depth reconstruction solely on the valid focus-measured signals that remain. Additionally, a straightforward calibration method for the posture angles of the vision-motion system is developed, ensuring that the reconstruction system produces aligned image sequences and is utilized alongside the proposed reconstruction method to achieve a high-quality depth map. Compared to conventional shape reconstruction methods that do not utilize the classification of focus-measured signals, the proposed method demonstrates significant advantages in reconstruction quality and accuracy. The results indicate that the proposed method achieves remarkable performance in signal classification, demonstrating an excellent ability to separate noise and minimize error in the depth map. This enables the generation of more accurate, high-quality depth maps. Moreover, the proposed method can learn and continuously improve its reconstruction performance through further training, effectively addressing the adverse effects of focus-measured signal variability on shape reconstruction. In summary, the proposed method not only creates high-quality depth maps for various precision measurements but also serves as the core technique for 3D digital microscopes.