Quality enhancement of interferometric fringe pattern based on deep-learning-based denoising of combined noise

Abstract

Optical interferometry is a highly accurate method for measuring the properties of optical components. This method uses fringe patterns obtained using an interferometer to extract the phase based on the properties of the measurement target. Noiseless fringe patterns are ideal for precise measurements. However, owing to unavoidable factors in real-world environments, noise is inevitably introduced into fringe patterns. Therefore, a denoising process is essential for enhancing the phase extraction results. Among the different types of noise, the second-harmonic component has been reported to be the most dominant. In this study, the combination of the second-harmonic component and Gaussian noise is defined as combined noise. In addition, a combined noise decomposition method using deep learning is proposed. Specifically, the denoising residual attention UNet (DRAUNet) model was designed and trained to decompose combined noise from fringe patterns. The model can uniformly increase the contrast of noisy fringe patterns. The proposed method was validated using simulation data, confirming its superior denoising performance compared with those of other methods. Furthermore, the necessity of considering the second-harmonic component in the denoising process was verified. Finally, the effectiveness and superiority of the developed method were further validated using real fringe patterns from the surface shape measurements of a silicon wafer obtained using a Fizeau interferometer.