| Color management is used to implement the transform between different color spaces,to achieve the color unity of cross-device or cross-platform,and solve the problem of color distortion in color reproduction process.In ICC-based color management,the device-independent color space is chosen as the Profile Connection Space(PCS).Color of different devices can be converted through PCS merely by establishing the relationship between device color space and PCS in advance,thus achieving the "WYSIWYG" purposes.ICC Profile provides the conversion relationship between device color space and PCS for the ICC color management system.This thesis aims to generate the ICC Profile of display device.It focuses on the study and innovation on the display characterization method,and designs an improved method for ICC Profile generation.The common methods of display characterization are reviewed and analyzed;these methods include polynomial regression,Tone/Matrix method,Multi-dimensional Look-up Table method,and artificial Neural Network.It summarizes that the characteristics of BP neural network are obtained,such as fuzziness,high efficient parallel processing and nonlinear mapping,which are consistent with the characteristics of the color space conversion process.In order to represent the conversion relationship between device color space and PCS more accurately,and solve the problems of artificial Neural Network,such as the low convergence rate and time-consuming,this thesis proposes an improved BP neural network for the generation of device characterization.In the algorithm design,Firstly,in view of the assignment problem of the initial weight vector in BP neural network,Steffensen accelerated iterative method is employed to accelerate the convergence of BP neural network.Secondly,conventional BP neural network converges slowly at the adjustment algorithm of weights and thresholds.To address this problem,this thesis introduces adaptive learning rate and the momentum term method to adjust weights and thresholds in the neural network.Finally,existing BP neural networks require high training cost.Considering that output Lab values are only affected by RGB without mutual influences of Lab channels when RGB is transformed to Lab,the proposed algorithm adopts the network decomposition to reduce input and output terminals,and so as to reduce the training scale of the BP neural network.After the algorithm of generating the ICC Profile has been improved,in order to solve the problem of loss of color information caused by without consideration of the characteristics of the specific image in the gamut mapping process,an improved ICC color management mechanism based on image self-adaptive gamut mapping is studied.The mechanism can automatically choose the suitable rendering intent.On the other hand,a color gamut mapping method based on image self-adaption is suggested.Under this mechanism,Contents recorded in the ICC Profile are designed,in order to achieve the best result of color reproduction.The algorithm,which is proposed in this thesis,is validated through real experiment by comparing the lab colour difference.In case of the same test sample sets,the Lab colour differences are calculated respectively by using the methods of the polynomial regression,the Tone/Matrix,the traditional BP neural network and the improved BP algorithm.Experimental results show that the display characterization method proposed by this thesis has a good performance on accuracy of color reproduction,and can effectively solve the problem of color distortion caused by the transform between different color spaces. |
