Research On Orthogonal Dictionary Learning Method For Visible Near-Infrared Spectral Reflectance Reconstruction

Spectral reflectance is of important value for target detection and identification,but existing spectral instruments such as interferometers are computationally intensive and have long spectral formation time,which makes it difficult to obtain the spectral reflectance of fast moving targets,and the existing spectral reconstruction algorithms have harsh experimental conditions and poor spectral reconstruction accuracy.This paper focuses on the method of introducing camera imaging characteristics as constraints in the process of spectral dictionary construction and spectral inversion,which improves the inversion accuracy and has important research value.First,this paper proposes a fast orthogonal basis dictionary learning algorithm to learn a sparse coding dictionary based on the data-set generated by spectral reflectance clustering,which effectively reduces the number of atoms in the dictionary.The algorithm also introduces an orthogonal constraint to turn the over-determined problem into a minimum solution problem,and updates the atoms in the sparse dictionary by hard threshold operations,which can shorten the algorithm training time while avoiding the similarity of atoms in the dictionary,and its reconstruction performance is improved.In the process of estimating the spectral reflectance,the spectral response curve of the visible camera needs to be obtained.The previous method first models the optical system of the visible camera and generates the spectral response curve based on the established model.However,this modeling process is too complicated,resulting in a low accuracy of the acquired spectral response curve.In order to improve the accuracy,the spectral response curve of the visible camera is obtained by actual measurement,and the measurement error is eliminated in terms of the stability of the optical power of the monochromator,etc.Meanwhile,in order to simplify the model and reduce the influence of the measurement error caused by noise,the measurement accuracy is improved by eliminating the background noise and optimizing the experimental operation and other steps.Finally,based on the trained spectral dictionary,along with the imaging principle of the camera,the mapping relationship between spectral reflectance and RGB pixels is established.For the input test image,white balance processing is first performed to normalize the scene illumination,and then the pixel values are converted to chromaticity coordinates to eliminate the effect of shadows,and the closest chromaticity coordinates of Euclidean distance are searched for,and the spectral reflectance reconstruction of the target pixels is completed by the orthogonal matching tracking algorithm,and the experimental results show that the reconstruction accuracy of the method in this paper is better than those of the comparative algorithms.