Research On Reconstruction Of Human Skin Reflectance Based On Digital Camera

With the development of society,people have higher and higher requirements for the reproduction accuracy of object color.Color values,such as RGB,XYZ or Lab have been difficult to accurately represent the color information of an object,and spectral reflectance,which can exactly represents the color of an object,is a property presenting the color after the light reflected from an object.In daily life,the spectral reflectance of human skin has been widely used in face recognition,animation rendering and cosmetic manufacturing,etc.,but there are few domestic studies on its reconstruction.Then,this thesis is going to make a research on the reconstruction algorithm of the human skin spectral reflectance.Traditionally,the spectrophotometer is used to measure the spectral reflectance of an object with the accurate obtaining result,but it has many drawbacks.For example,with its high price and contact measuring equipment,the spectrophotometer can only measure one color block at a time,and it requires that the surface of the test sample must be uniform.Recent years,the method of using the digital cameras to obtain the color information of an object makes up for the deficiency of spectrophotometer,but its measurement accuracy is lower than the traditional spectrophotometer,so how to accurately reconstruct spectral reflectance by using RGB response value of digital camera has become one of the research hotspots today.This thesis is devoted to reconstructing the spectral reflectance of human skin by using the RGB response value of digital camera.Based on the different input color information in the spectral reconstruction algorithm,the spectral reflectance is reconstructed from two angles of direct reconstruction(i.e.camera RGB response value)and indirect reconstruction(i.e.XYZ tristimulus values obtained through characterization model),and the reconstruction results are evaluated by root mean square error and color difference.As the characterization model between the RGB response value and the XYZ tristimulus values of the camera needs to be established in the indirect reconstruction process,in this thesis,the polynomial model method is used to obtain the characteristic matrix and the transformation accuracy of the common polynomial model and the root polynomial model is compared.The experimental results show that the polynomial model is related to the model order,the number of extended terms and training samples,and both models can give good performance in the 3rd and 4th order.In the direct reconstruction method,the camera sensitivity function information is unknown in many cases,so this thesis also compares the spectral reconstruction accuracy of different reconstruction algorithms when the camera sensitivity is known and unknown,and evaluates it by using root mean square error and color difference.The results show that in realscenes(with noise),Wiener estimation with known camera sensitivity performs best,followed by pseudo-inverse method and principal component analysis method with unknown camera sensitivity.In this thesis,the multiple light source technology combined with the traditional three-channel camera is used to simulate the imaging of the multi-channel camera.It is verified that the multi-channel camera composed of reasonable light sources can improve the spectral reconstruction accuracy,and it is also verified that the smaller sampling interval and higher image bits can obtain better spectral reconstruction accuracy.Standard color cards are often used in the spectral reconstruction,but the number of them is large and redundant,which greatly increases the computational work.In order to represent all spectral and color information with a small number of samples,a sample selection method based on color space is proposed in this thesis.First,the color difference between the original training sample and the test sample is calculated and arranged in ascending order.Next,the sample whose color difference is less than a certain limit value is selected as the optimized training sample set of the test sample.Then,the test sample is reconstructed by using this optimized training sample set.By analogy,the spectral reconstruction of each test sample can be obtained.Compared with the existing methods,the proposed sample optimization scheme has better root mean square error and color difference.