| Optical parameters such as absorption coefficient(?_a),scattering coefficient(?_s),anisotropy factor(g),and refractive index(n)are paramount to the research of physical structures and chemical compositions of media.However,current solutions to the inverse problem of optical property parameters heavily depend on the theoretical model.There is a huge identification error and the solution is often trapped into local minima.At the same time,the existing algorithms cannot estimate all the four parameters,in which only one or two parameters are studied.All of these greatly reduce the effectiveness of optical inspection technology.As an important feedback information in the field of control,optical reflective information,its accuracy affects the control system.The accuracy of optical parameters plays an important role in the control of biomedical treatment and food processing industry.When human cells become cancerous or abnormal precipitation of pigments in human skin,the tissue optical parameters of the corresponding parts will be different from normal tissue;when the content of a chemical component in food is higher or lower than the normal level,its optical parameters will also change accordingly.Therefore,it is meaningful to study optical parameter estimation techniques.The thesis includes development of simulation software for diffusive light intensity distribution image and the research of multiple parameters recognition method based on image optical characteristics.The main contents are listed as follows:1.Development of simulation software for diffusive light intensity distribution image generation.The traditional Monte Carlo simulation program can only provide the diffusive light intensity information of excited by the vertical incident light at the surface of the media,which is not suitable for the simulations of mass samples because of the low computation speed and inconvenient input of optical parameters.In order to enhance the richness of sample data,file reading and writing function was applied in this work to enhance the input and output mode of the traditional Monte Carlo program.On the other hand,Qt5.0transplanted the traditional Monte Carlo simulation program,in which the multi-threaded module speeds up the running speed of the program.It also provided a visual interactive interface,which greatly improved the convenience of the program to obtain spatial diffusion information.2.Research on a multiple-parameter identification method of optical characteristics based on multiple-distance response distribution of light intensities from media surface.In order to enhance the richness of the reflective data from signal response perspective and train the high-precision optical characteristic parameter recognition model,the diffusive light intensity distribution images at different distances from media surface were simulated by simulation software to replace the diffusive intensity from Monte Carlo program at a single distance.In the developed simulation software,the diffusive light intensity distribution images of various media were simulated.Then five basic feature vectors of the images were used as the input of a BP neural network to train the network.The outputs of the network are the four optical parameters of absorption coefficient?a,scattering coefficient?s,anisotropy factor g and refractive index n identified.The BP(Back propagation)neural network provided a high precision and efficient identification method for them.The traditional method based on the nonlinear least square fitting curve has 12%and 15%relative errors for the identification of absorption coefficient(?_a)and reduced scattering coefficient(?_s').Compared with it,the proposed method not only identified four optical parameters simultaneously,but also greatly improved the identification accuracy.When the signal-to-noise ratio is 40dB,the average relative error for the identification of absorption coefficient,scattering coefficient,anisotropic factor,and refractive index are only 8.5%,10.1%,2.3%,5.7%,respectively.The recognition efficiency of the neural network trained in this work is much higher than that of the traditional least square fitting based on the mechanism model,and it is not affected by local minimum issues.3.Research on the method of multiple-parameter identification of optical properties based on a multiple angle stimulated diffusive reflection intensity distribution.The macroscopic characteristics of the tested medium,such as shape and size,may affect setup optical inspection equipment.In view of this problem,this work proposed a convolution neural network algorithm based on multiple angles diffusive reflection intensity distribution images.The angle of incident light is adjusted according to the specific shape of the medium to facilitate the detection of samples and increase the richness of diffusive reflection light information from signal excitation.The developed simulation software simulated the diffusive light intensity distribution images of different angles of incident light to provide training and test sets for a Convolutional neural network to identify optical parameters.Finally,the average relative errors of a trained deep convolution neural network based on VGG16 for identifying absorption coefficient(?_a),scattering coefficient(?_s),anisotropy factor g,and refractive index n are 10.3%,12.0%,10.5%,8.2%,respectively.The results show that the network has excellent anti-noise performance.Parameter estimation accuracy and convergence are affected by the richness of data.Traditional optical parameter estimation method only use diffusive light intensity measured at the surface at the media,which highly reduces the richness of the data.This introduces the problem of not able to estimation the four parameters.This work enhances the richness of data and may estimate the four parameters at the same time.The application of neural network reduces parameter identification time and avoid local minima problem.This work is significant in the field of optical measurement. |
PDF Full Text Request