Research On Hyperspectral Remote Sensing Image Denoising Method And Its Application In Target Detection

Since the 1980s,thanks to the rapid development of satellite technology,sensor technology and computer technology,hyperspectral remote sensing technology has emerged and gradually become one of the important means of earth observation.Benefiting from its high spectral resolution,hyperspectral remote sensing is able to obtain continuous and refined spectral feature curves of objects,bringing favorable conditions for spatial feature analysis,and has been widely used in many fields such as homeland resource investigation,earth science and application and national defense security.The emergence of hyperspectral remote sensing technology has prompted the development of remote sensing applications in the direction of greater refinement and accuracy.However,because of many factors such as the internal and external environment of the imaging system,hyperspectral remote sensing images are inevitably contaminated by mixed noise,resulting in image quality degradation,which seriously hinders the subsequent information extraction and refinement applications.To address the above problems,this thesis takes hyperspectral remote sensing images as the research object,proposes efficient denoising methods based on the principle of low-rank representation and deep-learning theory,aims to solve the problems in the removal of mixed noise from hyperspectral images,and extends the research to target detection applications.The main research work includes the following:Firstly,starting from the mechanism of noise action in hyperspectral remote sensing images,we analyze different type of noise in images and their causes,introduce traditional denoising models and deep learning-based denoising models for hyperspectral images,and analyze the problems that still exist in these two types of models at present.The theoretical basis related to target detection is introduced,and the effect of complex noise scenes on the effectiveness of target detection applications is firstly investigated through simulation experiments,and the necessity of noise reduction processing before target detection is illustrated.Secondly,a denoising algorithm based on subspace representation coefficient regularization and deep prior is proposed to address the problem that traditional denoising methods based on prior constraints require explicit explicit prior and manual customization of image prior.This algorithm is based on compressed perception theory,and uses a plug-and-play framework to decouple the data term and regularization term of the denoising model based on low-rank representation,and implicitly defines the regularization term prior through a deep denoising network to solve the subproblem of subspace representation coefficient images denoising.This model combines the traditional priori constrained denoising methods and deep denoising networks.Synthetic and real data experiments show that this algorithm can effectively remove Gaussian additive noise and Poisson noise,while preserving more structure features.Thirdly,the denoising algorithm SPARCA-Net based on orthogonal subspace projection and residual channel attention mechanism is proposed for high-intensity and mixed noise removal in hyperspectral remote sensing images.First,an orthogonal subspace projection(OSPA)module is designed based on the spatial attention mechanism,which implements the orthogonal subspace projection process in the traditional method using deep networks.By projecting the feature map into its signal subspace,the noise can be separated from the main signal in the process of projection and reconstruction.The traditional convolutional denoising network can only use the local spatial correlation information of the feature maps,and this module can explore and characterize the local and global spatial correlation information of the feature maps at the same time.Then,the residual channel attention(RCA)module is designed based on channel attention,which can adaptively learn and characterize the interdependence between channels and explore the global correlation information of the spectral dimension of hyperspectral images.As a result,the present model can fully utilize the local and global spatial-spectral information of hyperspectral images.The experimental results show that SPARCA-Net can recover high-quality images in high-intensity,mixed-noise scenes.Fourth,in response to the hyperspectral remote sensing payload design and denoising algorithm selection requirements driven by target detection applications,a new analysis of the sensitivity of matching and anomaly detection efficiency to different noise scenarios such as Gaussian,stripes,deadlines,and impulse noise is carried out.The experimental results show that there is a difference in the impact of noise on target and anomaly detection effectiveness,the impact on anomaly detection effectiveness does not change with the scene and target size,and the noise intensity has a negative impact on detection effectiveness under different scenes;the impact of noise on target detection effectiveness depends on the scene and target size,and the noise intensity has a negative impact on detection effectiveness for targets with more occupied pixels and clear contours,and a negative impact on detection effectiveness for targets with fewer occupied pixels.For the point target that occupies fewer elements and is isolated,the noise intensity has a positive effect on the target detection performance.Finally,the experiment verifies that the denoising algorithm proposed in this paper improves the effectiveness of target detection applications in complex noise scenes.