Research On Hyperspectral Image Reconstruction Algorithm Based On Image Fusion And Spectral Super-resolutio

Hyperspectral imaging is an imaging technique that can collect and process information across the entire electromagnetic spectrum.Its most important characteristic is the combination of imaging and spectral detection techniques.With hyperspectral imaging,each spatial pixel is dispersed into dozens or even hundreds of narrow-band spectral bands,achieving continuous spectral coverage while imaging the target space features.This feature gives hyperspectral images(HSI)powerful spectral diagnostic capabilities,which have been widely used in remote sensing,agriculture,medicine,and other fields.However,in practical applications,hyperspectral imaging systems are often affected by limitations in incident energy.In the real imaging process,there is always a trade-off between spatial and spectral resolution.Therefore,obtaining reliable hyperspectral images with high resolution through reconstruction techniques is of great significance.This article mainly explores hyperspectral image reconstruction algorithms based on image fusion and spectral super-resolution,which include the following two parts:(1)Fusion reconstruction of hyperspectral images and multispectral images(MSI).This chapter proposes a method based on discrete wavelet transform and generative adversarial networks to fuse hyperspectral and multispectral images.The feature extraction part of the generator network extracts the features of hyperspectral images through wavelet transform downsampling and upsampling modules,obtaining both spectral and frequency domain information.This method can reduce the number of parameters and improve the efficiency of feature extraction.In the fusion reconstruction part,the multispectral image information of multiple scales is gradually fused during the upsampling process to restore the missing spatial information of hyperspectral images and reconstruct high-resolution hyperspectral images.The discriminator network makes the generated hyperspectral image as similar as possible to the ground truth image,improving the quality of the reconstructed image.In addition,the loss function adds focal frequency loss based on mean square error loss and adversarial loss,narrowing the frequency domain gap between the generated and real images.Qualitative and quantitative experiments show that this method performs excellently compared to existing deep learning methods.(2)Super-resolution reconstruction of single RGB image spectra.Ordinary cameras can capture RGB images at low cost and with ease,and there is a correlation between RGB pixels and hyperspectral radiance values.Therefore,reconstructing high-quality hyperspectral images from RGB images can significantly reduce the cost of obtaining hyperspectral images.This chapter proposes a self-supervised hyperspectral image reconstruction network based on spatial-spectral Transformer for reconstructing hyperspectral images from a single RGB image.By combining the local feature extraction capability of convolutional neural networks and the long-distance dependency capturing ability of Transformers,it can more effectively utilize local and global information.In addition,the non-local spatial self-attention and spectral self-attention simultaneously focus on the spatial similarity and spectral correlation of hyperspectral images.Finally,a self-supervised restoration network is constructed to assist in constraining the reconstruction of hyperspectral images.Qualitative and quantitative experiments show that this method performs well in reconstructing hyperspectral images.