Comparing and analyzing generative models to denoise speckled velocity signals from laser doppler vibrometer

Abstract

Noise removal is a critical aspect of signal processing; however, existing algorithms often exhibit limitations when applied to complex and highly corrupted signals. In particular, suppressing speckle noise from vibration velocity signals acquired through laser Doppler vibrometry (LDV) remains challenging due to high computational cost and limited denoising efficiency. These methods frequently struggle to capture underlying signal patterns, while maintaining the processing speed required for real-time applications. To address these issues, a deep learning-based generative model, bidirectional generative adversarial network (BiGAN), is proposed to remove higher noise trends. The performance of this learning-based frameworks is evaluated against existing generative denoising approaches reported in the literature. The generative model is trained on preprocessed noisy-clean signal pairs, successfully learned noise characteristics and error trends, enabling robust speckle denoising under high-noise conditions. A comprehensive off-device analysis further reveals that the BiGAN combined with a low-pass filter (BiGAN+LPF) achieves the better performance, improving the signal-to-noise ratio (SNR) upto 18.87 dB for high noise level signals. In an on-device evaluation, the BiGAN+LPF framework establishes itself as the suitable method for suppressing speckle noise under a high sampling rate of 1 and 2MHz, achieving an average Delta SNR improvement of 17.16 dB.