A laser speckle image denoising framework based on zero sample self-supervised learning was proposed to reduce speckle noise in large-range laser triangulation systems. Firstly, a dual parameter Poisson-Gaussian mixture noise model with Poisson gain α and Gaussian standard deviation σ was constructed to characterize the noise characteristics of laser speckle. Secondly, a lightweight U-Net was designed and trained to estimate mixed noise parameters. By the generalized Anscombe transform, the nonlinear mixed noise domain was transformed into a Gaussian noise domain, and a noise distribution space suitable for self-supervised learning was established. Then, the Gaussian noise domain laser speckle image was decomposed into two sub images by down-sampling strategy, which were input into a lightweight multi-scale feature extraction network composed of a two-layer Inception structure with asymmetric convolution. This architecture learns the mapping relationship between two noisy sub images to indirectly infer the intrinsic mapping between the observed values of noise pollution and potential clean signals, without the need for pre collected clean image references. Finally, the noise domain was transformed into the original signal domain via the inverse Anscombe transform, and a clean image was obtained. The experimental results showed that this method adaptively suppresses speckle noise of different intensities while preserving image details, achieving a peak signal-to-noise ratio improvement of 3.2-5.8dB compared to traditional methods.

Zhuojiang Nan is a Research Assistant with Shanghai Jiao Tong University in the School of Automation and Intelligent Sensing. His main research interests include high precision photoelectric detection technology and new intelligent sensor. He has published more than 20 papers in reputed journals and has been authorized 8 Chinese invention patents.