Polsar Despeckling Neural Network Based on Local Texture Filters
YuFan Cai, Josaphat Tetuko Sri Sumantyo · 2023
This paper extends the Convolutional Neural Network to a neural network composed of local filters to diversify its processing unit called Speckle-Network-X (SNet-X), which can better realize the despeckling task on polarimetric SAR images. Neurons in SNet-X are designed based on Linear Minimum Mean Squared Error Filter (LMMSE-Conv) and Diffusion Filter (Diffusion-Conv). The combination of the mathematical modelling filter and the deep learning method enables these traditional local filters to achieve multiple iterations and possess self-adaptive capabilities, which is conducive to mastering the distribution information of noise and the local texture feature simultaneously. This paper also proposes a joint-training method using the loss function to make different filter neurons learn from each other and help the model improve the despeckling efficiency. The superiority of our model has been verified on simulation data, spaceborne linear polarimetric SAR data and airborne circular polarimetric SAR data. It is demonstrated that this method can be applied to any polarimetric channel for despeckling. Moreover, considering the lack of reliable reference images for real SAR data, this paper introduces a segmentation method named TeacherNet (TN) to evaluate the despeckling results, which is closer to subjective evaluation than statistical indexes. All the code and pretrained weights related to this paper will be made public on this website: https://github.com/YuFan-Cai/SNet-X.