Multibranch Separable 3-D Convolutional Neural Network for Hyperspectral Image Denoising
Haitao Yin, Hao Chen · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2023
3D convolutional neural network (CNN) offers a great potential spatial-spectral representation for hyperspectral image (HSI), and has achieved promising HSI denoising performance. However, the current 3D CNNs still suffer from limited non-uniform multi-component spatial-spectral features encoding, and incur a high computation burden. To address these issues, we draw inspiration from the success of multi-branch and separable convolution perspectives, and propose a plug-and-play Multi-branch Separable 3D Convolution Block (MS3CB) with different spatial-spectral receptive fields. Specifically, MS3CB comprises mixed regular and non-regular separable 3D branches, in which the non-regular separable 3D branch attempts to provide a meaningful non-uniform spatial-spectral features extractor. The separable 3D convolution in MS3CB factorizes the standard 3D convolution into a 2D spatial convolution and a 1D spectral convolution, which not only reduces model size, but also decouples spatial and spectral features in HSI for more flexible spatial-spectral representation. Based on MS3CB, we develop a novel 3D U-Net for HSI Denoising, called HSDU-Net, which uses MS3CB as the basic building block instead of standard 3D convolution block. We empirically verify that our separable 3D convolution block reduces about 64.6% parameters and achieves a certain performance gain over standard 3D convolution. Extensive experiments further demonstrate that HSDU-Net surpasses several latest baselines on various synthetic and real noisy HSI datasets.