| Determination of turbid media optical properties(absorption coefficientμ_a,scattering coefficientμ_s,anisotropic coefficient g,and refractive index n)via a non-invasive method has wide applications in medical diagnosis,food safety inspection,and material property measurement.Biological tissue structure has complexity and diversity,and it is impossible to explore the internal structure of a single layer,so it is necessary to detect the optical properties of the next or more layers of tissue through the surface layer to determine whether the biological tissue has lesions.Therefore,it is necessary to extract the optical properties of more layers of tissues to better quantify the propagation characteristics of light in multi-layer tissues and more accurately realize the non-destructive and real-time detection of multi-layer tissues.Such non-invasive optical detection method is crucial for biomedical and food safety applications,and so on.However,previous research on determination of turbid media optical properties focuses on identifying absorption coefficient and reduced scattering coefficient with single-layer or two-layer media,while very little effort was paid to determine the optical properties(_aμ,μ_s,g,n)along with the thickness d of each layer of a layered medium.Therefore,in order to solve this problem,the research of this topic will use diffusive light intensity distribution information corresponding to excitation light of different angles,combined with machine learning algorithm to realize the identification method of optical properties and layer thickness of each layer in layered turbid media.The main contents are listed as follows:(1)Study on generation of diffusive reflection signal in turbid media.For existing Monte Carlo simulation of optical transmission processes,the incident light is generally treated as a point with zero diameter.But in actual production applications,not only the diameter of the beam exists,but also when incident parallel light beam enters media,it is difficult to achieve strict collimation,resulting in a certain degree of divergence angle.In order to make the simulation results close to reality,in combination with the size of the incident spot and divergence angle,this work deduces the formula of the position and direction of the photon when the incident light is projected onto the surface of the medium.Select a reasonable range of optical parameters,set a reasonable internal mesh division,and finally obtain the diffusive light intensity distribution information through Monte Carlo simulation.(2)Research on the method of optical property multiple-parameter identification of single-layer media.The absorption coefficient,scattering coefficient,anisotropic coefficient,and refractive index of single-layer media are difficult to obtain and identify.To solve this problem,this work proposes to use diffusive light signals excited from multiple angles to enhance data richness,which will bring more constraints to parameter estimation.Previous research has proved that the diffusive light intensity profiles with different excitation angles are linearly independent,so multiple angle excitations may be employed to enhance data richness.In addition,the validity of the theory is also verified in this work considering the existence of the incident light spot diameter and the divergence angle.On the basis of constructing residual neural network in this work,the absorption coefficient,scattering coefficient,anisotropic coefficient,and refractive index of turbid media can be recognized by training model.The method proposed in this work was verified by Monte Carlo method simulating diffusive light intensities under various conditions.The incident light spot diameter and the divergence angle were considered in the simulation process.When the incident light spot diameter and the divergence angle were considered,and it was verified that the model trained by sample data was more accurate for the inversion accuracy of media optical parameters.Meanwhile,different levels of noise were added to diffusive light intensities to improve the generalization ability and anti-noise performance of the network.The results show that the average relative errors of the method proposed in this work for the identification of the absorption coefficient,scattering coefficient,anisotropic coefficient,and refractive index of turbid media are 8.6%,4.6%,1.7%,and 0.9%,respectively.The effectiveness of the proposed method is verified by comparing the recognition results of optical properties by different models.(3)Identification of optical properties and layer thickness of two-layer turbid media.Previous research on determination of turbid media optical properties focuses on identifying absorption coefficient and reduced scattering coefficient with single-layer or two-layer media,while very little effort was paid to determine the optical properties(_aμ,μ_s,g,n)along with the thickness d of each layer of a layered medium.With the number of layers increases,the number of unknown parameters will increase proportionally,which brings difficulties to parameter estimation.Therefore,in this work,Monte Carlo simulation is used to obtain the diffusive light intensity distribution information reflected by light propagation in two-layer turbid media,and random noise is added to diffusive reflection data for data enhancement.The neural network structure was constructed,and the diffusive light intensity signal under multi-angle excitation was used as the input,and the corresponding optical parameters and layer thickness of each layer of a layered medium were used as the output to train and test the network.Monte Carlo simulations were employed to validate the proposed method.Results show that the method constructed in this work identifies the relative errors of optical properties(μ_a,μ_s,g,n)and of the thickness d of each layer of the two-layer turbid media along the photon incident direction can be as low as 8.33%,7.43%,3.01%,1.71%,15.20%;8.32%,7.42%,3.02%,1.72%,15.36%,respectively.Through different models,including BP neural network,convolution neural network and residual neural network,by comparing the inversion results of five parameters(μ_a,μ_s,g,n,d)in each layer of two-layer media with different algorithm models,it is found that the residual network has better recognition results of optical characteristic parameters,and can still realize the layer thickness estimation of layered media. |
PDF Full Text Request