MLMVN in Speckle Noise Filtering
Igor N. Aizenberg, Olivia Keohane, Alejandro Lara · 2020
In this paper, we use the multilayer neural network with multi-valued neurons (MLMVN) as an intelligent tool for speckle noise filtering. MLMVN is a complex-valued feedforward neural network, which was successfully used for solving various problems including classification, prediction and additive noise filtering. Here, we apply MLMVN containing a single hidden layer for speckle noise filtering by processing overlapping patches taken from a noisy image. A resulting image is obtained by averaging the resulting intensities over all overlapping pixels. To train MLMVN, a learning set created of randomly selected patches from various images is used. A network is trained to transform a noisy patch into a clear one. To make a learning process more efficient, an incremental approach is used. A learning set is divided into batches, and a network is trained by sequentially going over all of them. In terms of PSNR, this approach significantly outperforms Lee filter - traditionally used for speckle noise filtering, and the BM3D filter - commonly recognized as one of the best nonlinear filters ever.