Research On Spectral Reflectance Reconstruction Algorithm Based On Kernel Entropy Component Analysis

With the increasingly rich and colorful people's lives,the image color information acquisition requirements are getting higher and higher.In daily life,we can perceive a wide variety of colors and use digital cameras,color film and other input devices to reproduce the color you see.However,the color of the picture obtained through these input devices may have a certain visual differences in different environments.The spectral reflectance of the surface color of different objects is different,which is the objective property of the object and does not change with other factors such as observation conditions.Therefore,the color scientist proposed a spectral-based color reproduction technique to solve the metamerism phenomenon,which could reproduce the color under different conditions.The key of the technique is to obtain the spectral reflectance of surface of the object firstly.Then,we combined with the graphic method based on the principle of chromaticity and spectroscopy to reproduce the surface color information.Generally the object surface spectral reflectance can be extracted by spectrophotometer,multi-spectral camera,spectral scanner and other equipment,However,these methods are not widely used in people's daily lives because of the large area,non-contact measurement and expensive cost.In this paper,a broadband multi-spectral imaging system platform was built by ordinary commercial digital camera and conventional absorption filters.According to the principle of multi-spectral imaging,the spectral reflectance of the surface of the object can be reconstructed by the optimal spectral reflectance reconstruction algorithm.The spectral distribution of the surface of most objects in nature is continuous.So it can be represented by a linearly combination of several basis vectors.The main algorithm used to acquire the basis vector are principal component analysis,independent component analysis and singular value decomposition,which uses a linear dimensionality reduction method to extract the basis vector.In fact,there is a certain nonlinear components in the real data.In order to compensate for the shortcomings of the linear dimensionality reduction method,some scholars have combined the kernel function with the principle component analysis(PCA)to form the kernel principal componet analysis.The kernel principal componet analysis algorithm uses the kernel mapping method to analyze the data,which needs to choose a kernel function and parameters that have not yet formed a standard for their choice.Compared with the kernel principal component analysis,the kernel entropy component analysis have combined with the entropy information theory,which can improve the nonlinear processing ability of the data.In this paper,we mainly study the linear and non-linear methods of extracting the basis vector of the spectral data.The principal component analysis and kernel entropy component analysis are discussed.The spectral reflectance of the test sample under different algorithms is reconstructed through the method that is the polynomial model combined with different basis vectors extracted by different feature extraction algorithms.Finally,the reconstrcuted spectral data are evaluated by spectral and chromaticity error evaluation methods.In experiment,SG color cards are used for training samples,oil paintings are used for the test samples.The preliminary experimental results show that the spectral reconstruction algorithm based on kernel entropy component principal analysis is superior to principal component analysis and kernel principal component analysis in chromaticity accuracy and spectral accuracy.So it has a certain application value for the real reproduction of the surface color of the object.