Research On Coded Aperture Compressive Spectral Imaging Reconstruction Algorithm Based On Compressed Sensing Theory

Compared with the commonly used RGB three-color imaging,spectral imaging can obtain more spectral response of target objects,so it can obtain more features and details of the target,thereby widely used in various fields such as remote sensing,agriculture,and biology.Traditional spectral imaging technology usually uses scan-based methods,such as scanning,wire scanning,etc.,scan a single point or a line of spectral curve at a time,and then move the entire scene.These imaging methods are exchanged for spectral resolution with time resolution and cannot shoot dynamic scenes.In response to this problem,some researchers applied compressed sensing to spectral imaging,and proposed the coded aperture snapshot spectral imager(CASSI).It modulates the target scenario information through the encoding template and prism,obtains the projection of the spectral cube on the camera,and then reconstructs the target spectrum data by calculating.This method can obtain target spectral information through a exposure,and has the ability to shoot spectral videos.However,there are two major problems in CASSI:(1)Due to the sparse assumptions of the target scene,the reconstruction error is inevitable;(2)the calculation complexity of the reconstruction algorithm is high.These two disadvantages are important factors to limit the practical application of compressed spectral imaging.This article will be committed to solving these two major problems and reducing the reconstruction time while achieving high-quality reconstruction.First of all,this article introduces the development history of spectral imaging technology.This article introduces different types of spectral imaging technology,analyzes their respective advantages and disadvantages,and then focuses on the principle of compressed spectral imaging and research status.Then,in order to improve the reconstruction time while improving the quality of reconstruction,this article proposes a reconstruction algorithm based on sparsely expressed.First,the local structural changes in different bands of natural scene spectrum images,and the correlation between the structure correlation with the corresponding RGB image were studied.For images obtained by the RGB branch in the dual camera system,use clustering and principal component analysis to quickly train the dictionary,and use this image to guide the choice of the dictionary.Select a suitable dictionary for each image block of the target spectrum image.Through sparse indication,the image information of the two branches is fixed.Finally,an adaptive guided filtering method is designed to enhance the anti-noise ability of the algorithm.The experimental results show that this method can also have a relatively short reconstruction time while obtaining high-quality reconstruction.In order to further improve the quality of rebuilding,this article makes full use of the local and non-local similarity of spectral image space and spectral dimensions,and proposes a reconstruction algorithm based on nonlocal self-similarity.First,the image obtained by the dual camera system RGB branch is used to build a three-dimensional image block,and then use this three-dimensional image block to assist the dictionary to learn and similar estimates.Finally,the target spectrum image is reconstructed by constructing a sparse representation model.Experiments show that this method can improve the quality of reconstruction,PSNR values increased by 45%-60%compared to other methods,and the use of three-dimensional image patch assisted dictionary learning and similarity estimation can effectively save the reconstruction time.In order to further save the reconstruction time,this article proposes a non-iterative fusion algorithm.According to the structural changes of different bands of spectral images and the similarity of the structure of the corresponding RGB image,each patch of the spectrum image to be rebuilt is regarded as an image patch and the product of a coefficient of an RGB image.Then use the RGB image to build a spectral group,use the CASSI branch image estimation coefficient,and finally obtain the reconstructed spectral image through the product of the spectral base and the coefficient.This article proves that this method can greatly reduce the reconstruction time,the average reconstruction time is within 3 seconds,and the non-iterative fusion algorithm of this article can combine other spectral reconstruction algorithms.The fusion result of this method can be used as the starting point of the initialization of other algorithms to improve the efficiency of reconstruction.