Research On Hyperspectral Image Restoration Technique Based On Deep Learning

The hyperspectral image(HSI)captured by hyperspectral sensors is a 3-D data cube containing two spatial dimensions of pixels and one spectral dimension.In general,an HSI contains a high spectral resolution to describe the detailed spectrum reflected by different materials in a scene,so that it plays a significant role in a wide array of applications,such as classification,clustering,and detection.Nevertheless,because of the sensor restriction and atmospheric interference,HSIs often suffer from various types of noise,such as Gaussian noise,stripe noise and deadlines.In addition,due to the hardware limitations of the optical imaging system,HSI usually has poor spatial resolution and low signal-tonoise ratio than panchromatic image or multispectral image.These factors greatly limit the application of HSI in the follow-up tasks.Thus,it is essential to design effective denoising and super-resolution algorithms to restore hyperspectral images.For HSI denoising problem,we propose an attention-based deep residual network(DRAN)to directly learn a mapping from noisy HSI to the clean one.To jointly utilize the spatial-spectral information,the current band and its K adjacent bands are simultaneously exploited as the input.Then,we adopt convolution layer with different filter sizes to fuse the multi-scale feature,and use shortcut connection to incorporate the multi-level information for better noise removal.In addition,the channel attention mechanism is employed to make the network concentrate on the most relevant auxiliary information and features that are beneficial to the denoising process best.To ease the training procedure,we reconstruct the output through a residual mode rather than a straightforward prediction.Simulated experimental results on several datasets demonstrate that our proposed method outperforms the state-of-the-art methods both in quantitative and visual evaluations.Hyperspectral image restoration is an ill-posed problem in essence.Certain prior knowledge needs to be adopted to guide the solution process.With the help of variable splitting technique,we can decompose the solution process into two subproblems of iterative optimization.One of the subproblems is a quadratic regularized least square problem,which can be easily solved by back projection or conjugate gradient.The other sub problem can be regarded as a denoising process,so we can introduce the deep denoising model obtained in the previous step.By combining model-based optimizationmethod and learning-based method,we can flexibly deal with different inverse problems,such as hyperspectral image denoising and super-resolution.With this strategy,we can further improve the performance of hyperspectral image denoising,and achieve better results than the existing methods in hyperspectral image super-resolution task.