Research On Multispectral Image Reconstruction Technology Based On Convolutional Neural Network

Multispectral images contain rich spectral information.With the continuous development of the field of spectroscopy,increasing amounts of the requirement for multispectral images is need to satisfy.Currently,there are two ways to obtain multispectral images,one is to use a traditional imaging system,and the other is to use a spectral reconstruction algorithm to simulate.The former can accurately obtain the multispectral image of the object through the imaging system,but requires the use of expensive equipment and technicall personnel to manipulate.For the purpose of solving the defect of the former,some scholars have proposed to use spectral reconstruction algorithms to obtain multispectral images.Using this method to acquire multispectral images neither requires expensive equipment nor consumes a lot of time.Since spectral reconstruction is to solve the serious ill-conditioned problem of three-channel to multi-channel mapping,the reconstructed multispectral image has relatively large deviation.To overcome the drawbacks of low accuracy of spectral reconstruction algorithm,a reconstruction means that using convolutional neural network is dishsed.Convolutional neural network methods can be subdivided into supervised learning methods and unsupervised learning methods.Based on the existing dataset and the poor learning ability of the residual dense network for channel information,a supervised learning method—adaptive residual dense network is produced.This network treats the residual dense structure as the basic framework.A squeeze and excitation module is insert to enhance the network's ability to learn channel information.So as to test the validation of the raised method,four mainstream reconstruction algorithms are compared with the proposed supervised method in the experiment.The test results prove that the proposed supervised method is superior to the comparison method in terms of subjective feelings and objective indicators.The applied means have great impact on multispectral images reconstruction.Supervised methods need to use a large number of multispectral datasets for reconstruction,it is very difficult to obtain multispectral data in natural scenes.Secondly,the reconstruction results of supervised methods rely heavily on datasets.In order to deal these two problems,an unsupervised learning method is produced—a reconstruction method based on the self-encoder.The self-encoder uses an improved convolutional network based on the hybrid attention mechanism,whichreconstructs the input RGB image into a multispectral image,And then reduce the dimension of the multispectral image to RGB image through downsampling.In order to verify the effectiveness of the proposed unsupervised method,this method is compared with the typical supervised learning method and the typical unsupervised learning method in the experiment.The experimental results show that the reconstruction accuracy of the proposed unsupervised method is higher than that of the typical unsupervised learning method,and is comparable to the best accuracy in the supervised learning method,which proves the effectiveness of the proposed unsupervised method.Finally,the application of multispectral reconstruction method in target recognition is introduced.Given a scene,using spectral reconstruction technology to reconstruct it into a spectral image,and then preprocess the multispectral image to obtain the spectral reflectance of the target area,and compare it with the spectral material library,and the physical properties such as the target material are determined to realize target recognition.