Research And Implementation Of Multispectral Reconstruction From Single RGB

Multispectral images can enable people to efficiently identify the materials of a scene,which has a wide range of applications.A multispectral image consists of multiple spectral bands formed by discontinuous bands,and each spectral band is a gray image.The issue of this paper is the multispectral reconstruction from a single RGB,in which multispectral refers to the image of about 30 spectral channels.Compared with multispectral images,RGB contains less information due to the small number of spectra,so this is an ill-posed problem.At the same time,metamerism increases the difficulty of reconstruction.Traditional methods model using image features for the issue,but there are many problems,such as too many artificially set parameters,relying on the collection of training data,long-running time.The deep learning method lacks explicability and has generalization problems.To solve this problem,we proposed to reconstruct multispectral from single RGB based on dictionary atom embedding.To solve the problems of traditional multispectral image reconstruction algorithm,firstly,considering the smoothness of the spectral change of multispectral image,we use the Gaussian process to describe the spectral quantization.Combined with the Bayesian nonparametric model,we reduce the dependence of dictionary learning on a priori parameter setting.We take the output result of online dictionary learning as the initial value and further accurately learn the spectral features based on a good dictionary,to achieve better accuracy.Then,using the dictionary obtained by the Gaussian process as the overcomplete feature set for reconstruction,we directly select the neighborhood from the sample and calculate the mapping between high-dimensional image and low-dimensional image according to the neighborhood embedding idea.The accuracy of the dictionary improves the accuracy of neighborhood selection and image reconstruction.In addition,we can calculate the data processing and mapping in the neighborhood embedding algorithm offline,so that the running time of online reconstruction is shorter.The results of experiment on the public data set show that the algorithm proposed in this paper has better reconstruction accuracy than the traditional algorithm.Meantime,the results of experiment also show that the camera spectral sensitivity is an important factor affecting the accuracy of multispectral reconstruction.Aiming at this issue,we construct a constrained least-squares objective function to estimate spectral sensitivity based on the nonnegativity and smoothness of spectral sensitivity.We use the alternating operator multiplier method to solve the objective function,that is,to estimate the spectral sensitivity directly from the image data.The experimental results show that the estimated spectral sensitivity can improve the reconstruction accuracy by 10%.