Research On Multispectral Reconstruction Method Against Exposure Variation

The digital image obtained by traditional color imaging technology has the problem of metamerism,while the multispectral image has the color characteristics of the observed object.The spectral image recorded with spectral information can effectively reduce the metamerism existing in color collection.Commonly used spectrum acquisition methods include hardware-based and reconstruction-based methods.The spectrophotometer and spectroradiometer in the hardware acquisition equipment adopt single-point measurement,and the color of the measured object is required to be uniform.Spectral cameras use imaging measurements,but are expensive and have limited application flexibility.Multispectral reconstruction based on RGB images has the advantages of convenience,speed,and low cost,and has become a research hotspot in multispectral image acquisition.At present,spectral reconstruction algorithms can be divided into two categories: machine learning and deep learning.Machine learning methods include pseudo-inverse method,principal component analysis method and kernel algorithm,etc.Deep learning methods mainly include convolutional network and confrontation network.However,neither the machine learning method nor the deep learning method can resist the change of the exposure environment,that is,the model established at a certain exposure level cannot be directly reconstructed at another exposure level.Otherwise,the shape characteristic shift will appear in the reconstructed spectral curve.This problem restricts the application of spectral reconstruction to scenes with variable illumination intensity.This article takes this as the research point to discuss.In order to solve the problems mentioned above,this paper proposes solutions for antiexposure changes for machine learning methods and deep learning methods.Among them,the deep learning algorithm is one of the machine learning algorithms.The reconstruction method based on machine learning requires less computing power and data sets than the reconstruction method based on deep learning,but its accuracy is lower than that of the deep learning method.Machine Learning-based Reconstruction Method:This paper proposes an adaptive weighted spectral reconstruction method based on root polynomial expansion.It first uses the root polynomial to expand the sample RGB image data,expands the RGB data to multidimensional data,and then uses the pseudo-inverse method to establish a spectral reconstruction model.It is used to solve the problem that the existing model is sensitive to changes in exposure levels,and then constructs an adaptive weighting matrix in the spectrally invariant feature space to further improve the accuracy of spectral reconstruction.The model is established on the basis of standard color card data and mineral pigment color card data,and the sensitivity of exposure changes is analyzed on the objective evaluation index,and then the method in this paper is compared with the existing similar type of spectral reconstruction method.Finally,the influence of the weighting strategy on the method in this paper is discussed.Reconstruction method based on deep learning: This paper proposes a spectral reconstruction algorithm through exposure enhancement and attention mechanism.This method first performs random exposure enhancement on the input RGB image,and performs exposure correction on the reconstruction result,so that the network can learn exposure invariant features on the basis of the original model.This enables a smaller network structure to resist changes in the exposure environment,and then embeds a hybrid attention mechanism in the model structure to further improve the accuracy of the reconstructed spectrum.Based on the multispectral data set,the research compares the method in this paper with the existing advanced spectral reconstruction algorithms in terms of objective evaluation indicators,and finally discusses and analyzes the advantages and characteristics of the method in this paper.The experimental results show that the two types of optimization methods proposed in this paper for deep learning and machine learning can both achieve the purpose of combating exposure changes,and the spectral reconstruction results have obtained lower loss valueson multiple objective indicators.It provides method support for the application of spectral reconstruction technology to natural open scenes.