Research On Deep Learning Algorithm For Hyperspectral Image Reconstruction From RGB Images

Hyperspectral images include multiple optical spectrum bands,which are widely used in remote sensing,medical diagnosis,substance detection,food inspection and agricultural production.However,the traditional hyperspectral image acquisition equipment is expensive and complex in design,and it is difficult to guarantee the real-time processing of high-spatial resolution image generation under the condition of high spectral resolution,which leads to difficulties in the development of hyperspectral image application.In recent years,the reconstruction of hyperspectral images using RGB images has become a hot topic because of the simplicity and high spatial resolution of RGB images.With the continuous development of deep learning,great progress has been made in the direction of using deep learning method to reconstruct hyperspectral images.However,most deep learning algorithms only consider spatial features while ignoring channel features.Aiming at the problems of existing algorithms,this paper studies and realizes the reconstruction of RGB image to hyperspectral image from the perspectives of expanding the receptive field and using channel characteristics.A dense multi-scale pyramid based RGB image-to-hyperspectral image reconstruction algorithm is studied.The algorithm uses the cross-channel fusion receptive field module to improve the reconstruction algorithm based on the residual block,increases the receptive field in the network to facilitate the extraction of context information,and reduces the amount of network parameters by modifying the side connections.In view of the influence of receptive field on spatial feature extraction in deep learning,the dense multi-scale network is adopted to improve the reconstruction network based on residual blocks,and the receptive field in the network is increased,which is beneficial to the extraction of context information.The improved multi-scale pyramid network is used for multi-level feature fusion to obtain more comprehensive spatial feature information.Dense connecting blocks are used to deepen the network and further improve the capability of feature extraction.The pixel shuffle can reduce the checkerboard effect and make use of channel information effectively.The algorithm improves the basic network by using the densely connected pyramid module,which not only enlarges the receptive field,but also helps to extract the features.It can obtain good reconstruction effect under the premise of controlling the size of the network model.A reconstruction algorithm from RGB image to hyperspectral image based on the attention mechanism is studied.The mixed-domain attention mechanism is used to extract spatial domain and channel domain information to further improve the reconstruction performance.This algorithm is based on the residual block based RGB image to hyperspectral image reconstruction algorithm.and uses the improved hybrid domain attention mechanism module to process the up-sampling part of the spectral dimension,and obtain the hyperspectral prediction with spatial attention and channel attention.Process the feature map,use the pixel attention mechanism to improve the main network part,and enhance the space and channel feature extraction capabilities of the feature extraction part.And reduce the generation of gradient problems by constructing shortcut connections.The algorithm uses a variety of hybrid domain attention mechanism modules,which effectively improves the algorithm’s reconstruction ability.In this paper,root mean square error(RMSE)and relative root mean square error(r RMSE)are used as evaluation indexes.The two reconstruction algorithms were compared with a variety of RGB image reconstruction algorithms for hyperspectral images,and the simulation experiment was carried out with RGB-hyperspectral image data set.The experimental results show that the reconstruction results obtained by the two reconstruction algorithms studied in this paper are closer to the actual hyperspectral images.