HSI-DeNet: hyperspectral image restoration based on convolutional neural networks

**2018_TGRS_HSI-DeNet: Hyperspectral Image Restoration via Convolutional Neural Network ( Chinese science- Chang Yi)**

Good evening everyone, today we share a paper on hyperspectral image restoration using Convolutional Neural Network (CNN).In this paper, based on CNN, we introduce both residual learning, dilation convolution and multi-channel filtering to obtain HSI-DeNet network, which achieves the removal of mixed noise in HSI: random, structural stripes, dead lines, etc. In addition, in order to obtain high image quality and quantization metrics at the same time, the proposed network is introduced into the generative adversarial network to obtain the HSI-DeGAN network in this paper.

First, the network structure of HSI-DeNet is shown below.

Each of these Blocks consists of three parts: convolution layer, block normalization (BN) and activation function (ReLU), as follows.

The input to the HSI-DeNet network is a three-dimensional hyperspectral data, which is trained by the network to get its noise, and then a residual learning strategy is used to get the final clean image. Dilated convolution and multichannel filtering are also used in the network, using dilated convolution to expand the perceptual field without increasing the parameters, thus enabling the use of more local and non-local information, and multichannel filtering to enhance the representation of spectral information. In addition, the padding strategy is used in this paper so that the image size does not change during the training process, with the following parameters.

The optimization of the HSI-DeNet network is based on a pixel-by-pixel loss function that is prone to over-smoothing results. In order to exploit the prior information at the image level, this paper proposes a new network HSI-DeGAN based on Generative Adversarial Network (GAN), the main idea is to use HSI-DeNet as a generative network and use a discriminative network to discriminate the truth from the real data and the HSI-DeNet generated data, both are trained alternatively to finally achieve a balance between the two, the overall framework is as follows.

The structure and parameters of the discriminant network are listed in the following table.

The dataset used in this paper is from the dataset ICVL containing 201 scenes, whose size is 1392*1300*519, from which the authors extract 500 and 50 subgraphs of size 180*180*10 for training and testing data respectively, and chunk the training data with a block size of 4040 and a move step of 40, so that 16 training blocks can be obtained for each subgraph, and then augmented by 8 times, so the final number of training data is 500*16*8=64000.

** The learning framework is****matconvnet**** , the running environment is Thepersonal computer with MATLAB 2014a, one Titan X GPU, an Intel i7 CPU at 3.6GHz, and 32-GB memory.**

Finally, we look at the experimental results, which are visualized on the simulated data as

The quantitative indicators are.

The visual effect on the real data is.