Study On Spatial Frequency Domain Imaging And Detection Of Optical Properties Of Pear

Nondestructive detection technology plays an important role in controlling the quality and safety of food and agricultural products.Visible and near-infrared spectroscopy(Vis-NIRS)technique,as a relatively mature method,has the advantages such as fast,nondestructive,strong penetration and has the ability to obtain internal quality information,was widely used.However,the basis of traditional Vis-NIRS method is the chemometrics method based on Lambert-Beer law.Due to most biological tissues are highly scattering medium,and the propagation of light in biological tissues is a very complex process,the Lambert’s law can not accurately describe the propagation process.The spectrum acquired by Vis-NIRS technology is generally affected by the absorption and scattering of light by samples.It can lead to poor prediction accuracy and poor model robustness by using such spectrum information to calibrate models.Therefore,it is of great significance to study the propagation mechanism of light in agricultural products and the detection techniques for optical properties of agricultural products.(1)The study of light propagation mechanism is helpful to understand the problems and reasons in traditional Vis-NIRS method.(2)The study of light propagation mechanism is helpful to the development of related detection devices for agricultural products.(3)The measured optical parameters can be directly used in the prediction of the physicochemical indexes of agricultural products.At present,the main methods used to obtain the optical properties of agricultural products are integrating sphere,spatially resolved(SR)and time resolved(TR)methods.The transport models used in the calculation of optical parameters mainly includes inverse adding doubling(IAD)algorithm,diffusion approximation theory(DA)and Monte Carlo(MC)simulation.In this study,the IAD based integrating sphere technique was used to investigate the feasibility and significance of the study.This work focused on the study of the emerging spatial frequency domain imaging technique,the main research contents include system development,fast computing transport model and contour correction.The ultimate aim is to apply this method to the rapid and nondestructive detection of optical properties of pear.The main research contents and conclusions are as follows:Fabricated different types of phantoms,measured and calculated the reference values of the corresponding optical properties.The manufacturing methods of liquid phantoms;thin,flat and solid phantoms;thick,flat and solid phantoms;semi-spherical solid phantoms were proposed.The reference optical properties of these phantoms were measured or calculated.The stability of the liquid and solid phantoms were tested and compared,the results showed that solid phantom was more stable than liquid phantom.Solid phantoms were used in the stability testing and periodic accuracy validation of the integrating sphere and SFDI systems.When calibrating or validating the system and algorithm which need large quantities of phantoms,the liquid phantoms was used.Improved the integrating sphere system,verified the feasibility of the study.By improving the automatic,single integrating sphere system,the optical properties of agricultural products can be measured in the range of 400-1100 nm,then it was used to obtain the optical properties of pear slices.The partial least squares regression(PLSR)was selected to build the prediction models for soluble solid content(SSC)and firmness,the determination coefficients of validation(R2v)were 0.40 and 0.44 respectively.Investigated the significance of the study by verifyed the enhancement effect of optical properties on the robustness of model.Taken milks as samples,by special grouping,lead to different sample related factors between calibration and validation sets.Compared the prediction effects of models based on absorption coefficient(μa),reduced scattering coefficient(μ’s),transmittance(T)and reflectance(R)spectra by using PLSR and stepwise multivariate linear regression(SMLR)methods.The results indicated that the fat calibration model based onμ’s spectra is more robust than the models based on μa,T and R spectra,and the protein calibration model based on μa spectra is more robust than the models based on μ’s,T and R spectra.Preliminarily developed the spatial frequency domain imaging(SFDI)system,and applyed the diffuse approximation equation(DE)transport model to the calculation of μa and μ’s.A low cost SFDI system was developed and calibrated.The linearity calibration results showed that the correlation coefficients were greater than 0.97 at 6 wavelengths and 16 spatial frequencies(fx)except at 658、675 nm and at fx=1 mm-1,and the correlation coefficients were 0.8908 and 0.9908 respectively,which revealed the poor linearity of the system when the fx is larger than 1 mm-1.The calibration slopes and intercepts could also been obtained at the same time.The accuracy of the SFDI system and diffusion approximation transport model was calibrated by thick,flat and solid phantoms.The maximum relative errors of μa and μ’s were 3.92%and 2.39%respectively after the calibration,which indicated that the system and algorithm could achieve high accuracy after calibration.The calibrated SFDI system and DE model were applied to the obtain the μa,μ’s maps of bruised pears.Normal and bruised pears could been distinguished from each other by coefficient of variation of μ’s maps,the discriminant accuracies were 100%and 98.3%respectively.The minor bruised pears(with impact energy of 0.025 J)could been distinguished from other two kinds of bruised pears by using contrast coefficients(A,B),the discriminant accuracies were 90%and 85%respectively.Studied the inverse calculation algorithm for turbid media based on the data obtained by Monte Carlo(MC)simulation,and built transport models by using least squares support vector machine regression(LSSVR).In order to overcome the limitations of DE model which is commonly used in SFDI technique,an inverse algorithm by using LSSVR combined with genetic algorithm based on MC simulation data at two fx(2-fx LSSVR+GA)was proposed.One of the two fx was determined as 0 mm-1,the other optimal nonzero fx was determined as 1/4 mm-1 by comparing the diffuse reflectance(Rd)obtained by experiment and simulation.The detection speed was increased by 8 times compared with the previous method.The forward validation results of two forward models showed that the determination coefficients(R2)were all lager than 0.99 and the root mean square error of validation(RMSEV)were all less than 3e-4.Compared the calculation results of 60 liquid phantoms by respectively using proposed 2-fx LSSVR+GA method,fitting to DE model by using 16 fx data(16-fx DA)and fitting to DE model by using 2 fx data(2fx DA),this process is inverse validation.The results showed that when the ratio of μ’s to μa(r(μ’s/μa))is less than 50,higher accuracy could be obtained by 2-fx LSSVR+GA method.The 2-fx LSSVR+GA method was then applied to the calculation of μa,μ’s of normal and bruised(with three different impact energy)pears.The results showed that the acquired μ’s maps could clearly revealed the contrast between the normal and bruised region in bruised pears.The disadvantage of this method lies in the time-consuming calculation.Improved the SFDI system and proposed a fast algorithm for the detection of optical properties of turbid media based on the snapshot method combined with machine learning algorithm.The hardware of the SFDI system was further automated,and the matching software was redeveloped,to realize rapid and automatic detection at multi wavelengths in the visible region.Introduced the snapshot method to realize the collection of only single diffuse reflection image for per sample and at per wavelength.The snapshot demodulation method was validated by 390 liquid phantoms,and the optimal fx was determined as 1/3 mm-1.Proposed the LSSVR and artificial neural network(ANN)models based on the MC simulation data,to calculate the μa,μ’s of turbid media.Compared the computational accuracy and efficiency of ANN,LSSVR,LSSVR+GA and DA models,the results showed that when the transport coefficient(μ’t)less than 1 mm-1 or the r(μ’s/μa)less than 5,the DA model showed great error,however,other 3 methods showed high accuracies.In terms of computational efficiency,the LSSVR model could achieve the highest computational speed,it took about 40 s for every pear sample.Therefore,the LSSVR model was taken as a fast algorithm and was applied to the calculation of μa,μ’s of pear,furthermore,the obtained were used to predict relevant quality indicators.The calculation results of bruised pear(with three different impact energy)which were detected at different time after bruise,showed that the bruised pear degrees could be clearly distinguished by μa and μ’s maps,however,there is no obvious difference between the original images.It indicated that the separation of μa,μ’s was helpful for the detection of fresh and subtle bruise which is invisible to naked eye.Compared with other methods,the model proposed in this chapter enhanced the contrast between bruised and normal area in μa maps while reducing detection time.This method not only improves the efficiency of experiment and calculation,but also provide higher sensitivity to μa.The prediction of multi physicochemical indexes of pear was investigated.The performance of the SSC calibration model based on μa at 6 wavelength by using ANN is not very well,the R2v was 0.468.However,the texture calibration models based on μ’s showed good accuracies,especially for brittleness and adhesiveness,the R2v were 0.839 and 0.865 respectively.Studied the nondestructive detection method by combining the phase measuring profilometry(PMP)with SFDI technique.The procedure and algorithm for the acquisition of sample surface height and tilt angle were established.The effect of the surface height on the μa,μ’s which calculated by commonly used DA model and rapid LSSVR model respectively was investigated by detecting the semi-spherical solid phantoms.The results indicated that for DA model,the μ’s showed the same trend as height,when the height of sample surface increased,the detected μ’s increased,and the μa showed the opposite trend.For LSSVR model,both the μa and μ’s showed the same trend as height,when the height of sample surface increased,both the detected μa and μ’s increased.MC simulation was carried on spherical sample,and the correction procedure and formula form of diffuse reflectance by surface profile information was determined.That is the diffuse reflectance is firstly corrected by the third-order attenuation exponential function with the height information,then the height corrected diffuse reflectance was further corrected by Lambertian correction.The proposed correction method was validated by semi-spherical solid phantoms,the results showed that relative errors between reference and calculated μa,μ’s values which was corrected by height and angle were small.The relative errors for μa,ranged from 0.78%to 11.84%,and for μ’s,they ranged from 2.83%to 8.95%.What’s more,the final μa,μ’s values after height and angle correction showed smaller standard deviations in region of interest(ROI)compared to the initial uncorrected values.It indicated that the height and angle correction makes the calculated μa,μ’s more uniform.The proposed method which combined the PMP with SFDI was applied to the detection of optical properties of pear sample.The results showed that for normal pear,the μa,μ’s maps were more homogeneous.For bruised pears,the location of bruised region could be highlighted,and the normal region became more uniform.