Deep Learning Methods For Hyperspectral Remote Sensing Images Stripe Noise Removal

Hyperspectral remote sensing images have numerous applications and play an important role in providing support in military target reconnaissance,precision agriculture,ecological protection and resource exploration.Along with the booming development of hyperspectral earth observation technology,hyperspectral images have shown an explosive growth trend.However,due to the influence of various factors in the imaging process and the imaging mechanism of hyperspectral images,the hyperspectral images are highly susceptible to random noise,strip noise and bad line pollution,and the real ground feature information is masked and lost,which seriously affects the subsequent processing of hyperspectral data and restricts its subsequent application.Therefore,the suppression and removal of hyperspectral image noise plays a crucial role in the application of hyperspectral data.Among the many noise factors,strip noise belongs to a special category of noise type,which has obvious directional and geometric characteristics and is more destructive to the structure of the image itself.The study of random noise removal is more common,but the special strip noise removal for the topic of the study is less common,considering the characteristics of the strip noise itself,targeted research strip noise removal has important value and significance.Traditional stripe noise removal methods often require a finely designed mathematical model to model the stripe noise and achieve the purpose of stripe noise removal by setting the optimal parameters,which is often difficult to select the optimal parameters and optimize the model solution,and the robustness of the parameters is not good enough to adapt to complex situations,and the use of the situation is limited.With the rise of deep learning methods in remote sensing image processing,it provides new ideas and technical support for hyperspectral image stripe noise removal,using its powerful modeling capability for generic features and semantic features of different levels to design hyperspectral image stripe noise removal methods based on deep learning.However,hyperspectral remote sensing stripe noise removal based on deep learning is still in its initial stage,and there is room for further research on the fidelity of the spectral and structural information of the restored images,and in addition,the extreme dependence of such methods on a large number of simulated training samples limits their application in real scenarios.In this paper,we address the two problems of spectral and structural information retention and improving model generalization ability in the framework of selfsupervised learning.Firstly,frequency prior and inter-band correlation are used to solve the problem of spectral structure information retention.Secondly,the model migration generalization ability is improved in terms of both the self-supervised training sample construction method and the combination of semi-supervised training.The main research contents are summarized as follows:(1)Based on a self-supervised learning framework,a frequency prior-driven selfsupervised network for strip noise removal from hyperspectral images is proposed for the retention of spectral and structural information.The network uses simulated strip noise with clean images to construct training samples,and utilizes the a priori knowledge that strip components and non-strip components in images behave as different frequency signals to preserve non-strip components with low frequency signals to ensure the minimum loss of spectral information when performing image recovery.At the same time,considering the spatial distribution characteristics of the strip noise and the strong correlation between the hyperspectral spectral bands,an attention mechanism of spatial spectral information enhancement is proposed to assist the network to focus on and learn the key features,so that the network can find a balance between strip noise removal and spectral information retention,avoiding the problem of focusing on strip noise removal and ignoring the restoration image quality,and enabling the network to have The network is able to improve the stripe noise removal effect by the ability of "where to look" and "what to look at"(2)Based on a self-supervised learning framework,a hierarchical feature cascade strip noise removal network with multi-modal strip noise collaboration is proposed.The network solves the problem of limited model generalization ability due to the training data made by simulated strip noise,and proposes a strategy to build a training sample set by using a large number of strip noise derived from images containing strip noise to break the single strip noise simulation method and break the limitation of insufficient model generalization ability due to simulated strip noise.On the other hand,the input data is converted into wavelet sub-bands after wavelet transform,which decomposes the data explicitly,reduces the difficulty of model training,uses the orthogonal property of wavelet transform to reduce information loss,and provides richer features for stripe noise removal by mining and integrating different levels of network features,and finally achieves the dual improvement of model stripe removal ability and model generalization ability.(3)A global-local modeling hyperspectral image strip noise semi-supervised removal network is proposed.The network integrates semi-supervised learning and self-supervised learning to further improve the model generalization capability using real strip-noise-containing images,realizes the direct use of unlabeled real data for network training,breaks the domain difference between simulated strip-noise and real strip-noise,and closes the gap between the model modeling of real strip-noise features.Explore the solution to the training problem of real stripe-noise containing data by unbiased estimation.At the same time,multivariate feature extraction modeling is used to integrate image global and local features,fully exploit long-distance dependent features and local spatial context information,strengthen the model’s ability to express complex unknown features,promote the model’s ability to cope with real scene strip noise removal,and enhance the generalization and robustness of the model.