Super-resolution Method Of Hyperspectral Images Based On Deep Spectral Constraints

Hyperspectral images can acquire continuous spectral information in the tens to hundreds of bands,while retaining the spatial and spectral information of the object being photographed.The continuous spectral information of hyperspectral images can directly reflect the material characteristics,and is widely used in mineral exploration,agricultural census,military target detection and other fields.However,Constrained by the existing sensor accuracy and cost issues,the images obtained by hyperspectral cameras restrict each other in spatial resolution and spectral resolution.Existing hyperspectral images have high spectral resolution but low spatial resolution.In practical applications,it is still difficult to obtain hyperspectral images with high spatial and spectral resolutions.This difficulty indirectly affects the implementation accuracy and research progress of detection,classification,and demixing of hyperspectral images.Therefore,improving the spatial resolution of hyperspectral images has gradually become a research trend while attracting widespread attention from researchers.Existing deep learning-based hyperspectral image super-resolution methods such as 3D-FCNN and SDCNN reconstructed hyperspectral images have spectral distortion problems,and interpolation algorithms need to be used in the preprocessing process to improve the spatial resolution of hyperspectral images.To improve the spatial resolution of hyperspectral images,the processes are cumbersome and the computational complexity is high,failing to maximize the advantages of deep learning technology in the field of image generation.In view of the above problems,this paper starts from two levels of network structure and loss function respectively,and proposes a hyperspectral image super-resolution algorithm combining channel attention mechanism and auxiliary classifier.The innovative work of this thesis mainly includes the following two parts:1.Research on hyper-resolution algorithm of hyperspectral image based on spectral constraint and attention mechanism.In the field of natural image super-resolution,the proposal of generative adversarial network(GAN)greatly improves the retention quality of the reconstructed image on detailed texture information.In the design of the network structure,this algorithm uses a generative adversarial network.By introducing a 3D convolution layer and a channel attention mechanism,the network can simultaneously effectively extract the spatial information and inter-spectral information of the input hyperspectral image.Setting a specific spatial constraint loss function and spectral constraint loss function can improve the spatial resolution of the hyperspectral image to be super-divided while suppressing the spectral distortion of the reconstructed image.This algorithm does not require the interpolation algorithm to preprocess the input hyperspectral image.It is an end-to-end hyperspectral image super-resolution network.2.Research on the super-resolution algorithm of hyperspectral image constrained by depth spectrum based on auxiliary classifier.The essence of the existing neural network adopting the attention mechanism is to selectively calibrate the feature map output by the middle layer of the network by generating a set of weight coefficients.This algorithm proposes a method for generating attention mechanism weight coefficients based on channel statistical variance values,which is used to improve the feature extraction and generalization capabilities of super-resolution networks.Based on the above research on hyperspectral image super-resolution algorithm based on spectral constraints and attention mechanism,this algorithm combines the auxiliary classifier idea and attention mechanism in the network structure,making the network model efficient in training and inference Adjust the value of the output feature map at the position corresponding to the transformation.Then,on the basis of the original loss function,the feature extraction ability of the network is enhanced by setting the auxiliary classification term loss function.Experimental data shows that,compared with existing hyperspectral image super-resolution methods,under the conditions of 2x and 4x reconstruction,the hyperspectral image super-resolution algorithm combined with channel attention mechanism and auxiliary classifier proposed in this paper can While effectively improving the spatial resolution of hyperspectral images while significantly suppressing spectral distortion,it has certain application value.