Research On Hyperspectral Image Reconstruction Algorithm Based On Sparse Adaptive Filtering

Hyperspectral images not only have properties similar to grayscale images,but also add one-dimensional spectral information,which brings great convenience for people to distinguish the composition structure and physical characteristics of the target,making it useful in aerospace,geography,medicine,etc.The field has been widely used.However,the amount of hyperspectral image data is huge.Compared with ordinary grayscale images,not only the resolution of the image is improved,but also one-dimensional spectral information is added.This makes the subsequent processing of the image very difficult,which also limits The development of hyperspectral images.The compressed sensing theory proposed in recent years has solved the above-mentioned problems in a different way from traditional sampling.In this paper,compressed sensing technology is applied to the reconstruction of hyperspectral images,and the reconstruction algorithm based on sparse adaptive filtering is studied.First of all,in view of the deficiencies of the SL0 series of algorithms in reconstruction accuracy and reconstruction speed,this paper uses a smooth function norm closer to the L0 norm to improve the accuracy of image reconstruction,constructing a composite sine function model,And combined with the re-weighted function on this basis,which promotes the sparsity of the signal and speeds up the convergence speed.The algorithm also introduces a regularization mechanism to transform the model into a linear programming problem,which enhances the robustness of algorithm in Gaussian noise.Next,for the "sawtooth phenomenon" of the steepest descent method and the sensitivity of the Newton method to the initial value,this paper adopts a new joint optimization method to solve iteratively.Based on this,the RCSFSL0 algorithm is proposed,which further improves the image reconstruction performance.Secondly,the adaptive filtering framework is applied to the compressed sensing reconstruction algorithm.Aiming at the problem that the traditional adaptive filtering algorithm degrades or even fails in the case of non-Gaussian noise,this paper uses the minimum error entropy(MEE)criterion to replace the traditional least mean square(LMS)constraint condition,because the error performance surface of MEE is much smoother than LMS algorithm,the MEE criterion is more suitable for non-Gaussian noise.On this basis,combined with the previously proposed Composite Sine function model,adding sparsity constraints,the CSFSL0-MEE algorithm is proposed,which improves its image reconstruction performance under mixed Gaussian noise.Finally,taking advantage of the strong correlation between the spectra of hyperspectral images,and combining the idea of variable sampling,a CSFSL0-MEE hyperspectral reconstruction algorithm based on variable sampling is proposed.This method first groups the hyperspectral images according to the correlation between the spectra.And sample the reference images in the group with a higher sampling rate,sample the non-reference images with a lower sampling rate,and then use the inter-spectral correlation to reconstruct the non-reference images according to the reference images,with a lower sampling rate average the sampling rate to reconstruct the entire hyperspectral image.Compared with the method of constant sampling rate,the reconstruction accuracy and speed of hyperspectral images are improved.