Self-Denoising of BOTDA Using Deep Convolutional Neural Networks

We propose the self-denoising network (SDNet), a self-supervised network based on a convolutional neural network (CNN), for Brillouin trace denoising. With the target noisy image as the only input, the proposed method has no hardware restriction, requirement for image priors, or assumption for noise distribution. The Bernoulli mask and the partial convolutional layer implemented in the encoding process help capture the input features efficiently, while the dropout in the decoding process extends the feasibility and generality of the proposed network. Experimental results indicate that for Brillouin traces with frequency steps of 0.5 MHz/1 MHz, the proposed SDNet can improve the accuracy of Brillouin frequency shift (BFS) estimation by 40%/61%, 32%/31% and 24%/32% under input signal-to-noise ratios (SNR) of 5.5 dB, 8.5 dB, and 11.5 dB, respectively, without degradation in spatial resolution. Furthermore, the SDNet demonstrates a robust denoising performance in BOTDA systems with different application scenarios.