Paper detail

DeepBessel: deep learning-based full-field vibration profilometry using single-shot time-averaged interference microscopy

Full-field vibration profilometry is essential for dynamic characterizing microelectromechanical systems (MEMS/MOEMS). Time-averaged interferometry (TAI) encodes spatial information about MEMS or MOEMS vibration amplitude in the interferogram's amplitude modulation using Bessel function (besselogram). Classical approaches for interferogram analysis are specialized for cosine function fringe patterns and therefore introduce reconstruction errors for besselogram decoding. This paper presents the DeepBessel: a deep learning-based approach for single-shot TAI interferogram analysis. A convolutional neural network (CNN) was trained using synthetic data, where the input consisted of besselograms, and the output corresponded to the underlying vibration amplitude distribution. Numerical validation and experimental testing demonstrated that DeepBessel significantly reduces reconstruction errors compared to the state-of-the-art approaches, e.g., Hilbert Spiral Transform (HST) method. The proposed network effectively mitigates errors caused by the mismatch between the Bessel and cosine functions. The results indicate that deep learning can improve the accuracy of full-field vibration measurements, offering new possibilities for optical metrology in MEMS or MOEMS applications.