Hyperspectral Snapshot Compression And Reconstruction Method Based On Tensor Decomposition And Depth Prio

Hyperspectral images are three-dimensional images composed of adjacent spectral bands,that provide a wealth of spatial and spectral information and are widely used for target identification and material detection.However,the storage and transmission of these large datasets can pose significant challenges.As such,there is an urgent need to develop effective strategies for obtaining high-quality hyperspectral images with reduced data requirements.To address this issue,compressed spectral imaging based on compressed sensing theory has emerged as a promising solution.This technique requires only a small amount of 2D measurement information to be acquired,allowing for efficient hyperspectral image reconstruction through the use of reconstruction algorithms.Despite their potential advantages,traditional compressive reconstruction algorithms often suffer from long reconstruction times and suboptimal reconstruction quality,which severely limit their practical utility.To overcome these challenges,in this thesis,we investigate a novel class of compressive reconstruction methods for hyperspectral images that leverage Tucker decomposition and exploit the full depth of image structure information as prior knowledge.The main work of this paper is as follows:In response to the shortcomings of traditional iterative optimization methods for hyperspectral image reconstruction,such as long reconstruction time and poor reconstruction quality,as well as the underutilization of image structure information in deep learning-based approaches,we have proposed a novel hyperspectral image reconstruction algorithm that leverages deep Tucker decomposition and spatial-spectral learning networks.Our algorithm integrates tensor decomposition to develop a Tucker decomposition model that learns low-rank a priori features of the image,and enhances feature representation by cascading diverse levels of features.Moreover,a spatial-spectral learning network,combined with a U-Net architecture,learns multi-scale spatial and spectral correlations to further enhance the image reconstruction quality.Experimental results show that our algorithm achieves impressive time efficiency and reconstruction quality on both simulated and real data,thereby enabling real-time acquisition and reconstruction in spectral imaging systems.In light of the issue of poor interpretability associated with end-to-end neural networks,we propose a novel hyperspectral image reconstruction algorithm that combines traditional iterative optimization methods with Tucker decomposition module,leveraging a joint Tucker decomposition-spectral Transformer deep unfolding network to enhance interpretability and image reconstruction quality.Our algorithm incorporates the Tucker decomposition module into the Transformer to create a joint Tucker-spectral Transformer module that captures both lowrank prior and global spectral information of the image.This module is also utilized to construct a U-Net architecture,which serves as an approximation operator for the solution during iterative optimization.Reconstruction is achieved by unfolding the cascade of all iterative optimization stages into an end-to-end trained neural network.Experimental results show that the proposed algorithm not only guarantees reconstruction quality,but also improves interpretability in the task of hyperspectral image compression and reconstruction task.