Research On Reflectance Estimation Algorithm Based On Locally Weighted Regression Method

Color is an important visual element and closely related to people's life and work.The collection and representation of color is playing an important role in printing,dyeing and imaging.A large amount of color gets into people's eyes with the rapid growth of the digitalization and informationization.People can see colorful scene and observe various material through screens and videos rather than coming in person.All of these require accurate color reproduction.Spectral reflectance is the intrinsic characteristic of object's surface and the most accurate and informative way to represent a natural object's color.With the spectral reflectance,we can reproduce the color accurately under specific illuminant and observer condition.Generally,spectral reflectance can be measured by professional measuring instrument such as the spectrophotometer.However,the spectrophotometer is expensive and work only in low resolution.Also the measuring speed of spectrophotometer is not applicable in real circumstances.RGB tristimulus values,on the other hand,can be easily collected by the widely used digital cameras and scanners.Thus the reflectance estimation methods from RGB responses have received increasing attention.This article studies mainly on the characteristic of the spectral reflectance and the spectral reflectance estimation methods from RGB tristimulus values.As the RGB responses can be seen as the sampling from spectral reflectance in the low dimension,the mapping from RGB responses to spectral reflectance is not unique.Because of the feature of smoothness,natural objects' reflectance resides on a lower dimensional sub-manifold which is embedded in the high-dimensional ambient Euclidean space.This characteristic ensures the effectiveness of many estimation methods in the low-dimensional reconstruction of reflectance.The classical Wiener estimation methods and finite-dimensional methods need to know the prior information of the imaging environment and the imaging system,which limit the application in practice.The regression methods need not to know the prior information thus are more practicable.Since the training samples are always insufficient in real circumstances,the distribution of reflectance between the training samples and the true reflectance would be different.Global regression methods are prone to over-fitting in this case.With the good generalization capability in small samples,the local linear regression method has caught a great deal of attention.But the local linear regression method is prone to under-fitting due to the rough selection and exploitation of the local samples.This article focus on the issues mentioned above.Firstly,a spectral reflectance estimation method based on locally weighted regression is proposed.This method gives different weights to each training sample through proper distance metric and bandwidth of weighted kernel,so it can exploit the local manifold structure well,alleviate the over-fitting and under-fitting and reconstruct the spectral reflectance more accurately.Secondly,local sample selection methods based on adaptive neighbor bandwidth and improved distance metric are proposed to improve the performance when the samples are unevenly distributed,make use of the color vector shape feature to improve the stability of the model and the efficiency of local samples.