Research On Hyperspectral Imaging Algorithm In Deep Learning

Hyperspectral images contain richer color information than RGB images,and have important applications in many fields such as military,agriculture,and food safety.Traditional spectral imaging systems have the characteristics of high cost,large size,and high complexity.With the development of computational photography,computational spectral imaging methods have made considerable progress.Among them,the spectral reconstruction algorithm based on compressed sensing is the current mainstream reconstruction algorithm.However,spectral reconstruction algorithms based on compressed sensing not only rely on hand-designed feature priors,but also need to encode imaging,resulting in complex systems,low luminous flux,difficult to calibrate,and inability to shoot dynamic spectral videos,hindering the popularization of spectral video imaging technology.In response to the above problems,this paper proposes two de-dispersive spectral reconstruction algorithms from the perspective of data-driven and model-driven,and builds a corresponding spectral imaging system based on this.The main work and innovations of this paper include:1.Build a compact spectral imaging system without coding,which consists of only a dispersive element（dispersive grating）and an image sensor.Compared with traditional compressed sensing imaging technology,the system does not require complex coding apertures,has a simple structure,and is more conducive to integration.2.Propose a data-driven de-dispersive spectral reconstruction algorithm.The algorithm uses a three-dimensional convolutional neural network to learn the spectral information from the dispersive RGB image and reconstruct the hyperspectral image.The model obtained by network training can directly de-disperse the dispersive RGB image outside the training dataset,realize real-time spectral reconstruction,and can be applied to spectral video imaging.3.Propose a model-driven de-dispersive spectral reconstruction algorithm.An unsupervised network is used to reconstruct a hyperspectral image from a single dispersive RGB image by online training.This method does not require a large number of datasets for network training,but builds a network based on a physical model,and obtains network weight parameters through online training and is used to solve underdetermined inverse problems.The algorithm is simpler and the data collection cost is lower.In summary,this paper constructs a compact and concise spectral imaging system,which gets rid of the complex optical structure and calibration process of traditional compressed sensing systems.Based on this system,this paper proposes two dedispersive spectral reconstruction algorithms,which use neural networks to learn spectral information from dispersive RGB images,getting rid of the prior limitations of manual feature design in traditional methods,making the algorithm more robust.Comparing these two spectral reconstruction algorithms and two traditional compressed spectral reconstruction algorithms,the PSNR and SSIM of the reconstructed image are improved by 1.15 d B and 3.24% respectively.