| Terahertz technology is considered as one of the foremost technologies in the 21st century and has caught widespread attention from scholars and industry.Located between millimeter wave and infrared,the terahertz frequency ranges from 0.1 to 10 THz with unique advantages such as safety,penetration,and fingerprint spectrum.It is broadly used in a variety of applications,including internal structure detection,safety inspection,spectral recognition,non-destructive detection,biomedical applications,and the like.A terahertz time-domain spectroscopy（THz-TDS）imaging system could have a bandwidth of several THz and can be used to obtain ample spectral information of the target through coherent detection,which is viewed as one of the most promising terahertz techniques.Currently,the THz-TDS technique faces some challenges,including a time-consuming data acquisition process,difficult to be applied for detecting multi-layer structured objects,low recognition rate with a single feature,limited spatial resolution determined by the beamwidth,etc.Based on a self-built experimental THz-TDS system,this dissertation mainly studies the terahertz imaging techniques theoretically and experimentally.The main accomplishments and contributions are summarized as follows:Firstly,we set up an experimental THz-TDS system with a 1560 nm laser and photoconductive switch,which can generate a terahertz pulse with a bandwidth of 4 THz.Then,we discuss the method to extract the feature parameters of samples for the THz-TDS,including refractive index,absorption index,dielectric constant,and material factor.When using the THz-TDS system to measure characteristic parameters,thickness measurement error has a significant influence.Therefore,we apply a genetic algorithm to improve the estimation of the sample thickness.In the case of classification,we proposed a method by combining multiple features such as absorption,refractive index,and transfer function as input to improve the classification accuracy.Three machine learning methods,Back Propagation neural network,Learning Vector Quantization,and Support Vector Machine,were applied to perform multi-feature classification and recognition.The experimental results show that the average accuracy is up to 95%,which means that the proposed method can effectively identify a variety of substances.Secondly,we introduce a maximum likelihood maximum expectation（MLEM）algorithm with a terahertz beam model to two-dimensional（2D）scanning imaging.The beam model for the terahertz time-domain spectroscopy system is a Gaussian function,which limits the spatial resolution of the image.A Wiener deconvolution is applied to reach the super-resolution for the 2D scanned images.Considering the Gibb effects in the Wiener deconvolution,we introduce an MLEM algorithm combined with a terahertz beam model into the 2D scanning imaging,which can effectively suppress the Gibbs effect while improving the image resolution.Experimental results have verified that this method can improve spatial resolution and suppress side lobes.Thirdly,we propose a robust soft-thresholding sparse recovery algorithm to mitigate the time-consuming process of THz-TDS data acquisition.The number of sampling data and sampling time are reduced through sparsely sampling in the image domain,and the image recovery is completed by an iterative method using interleaved sparse constraint in the wavelet domain and sampling data constraint in the image domain.Furthermore,a reweighted L1 norm and sparsity-average algorithm are applied to enhance the algorithm robustness.Experimental results show that the relative error of the recovered image is only about 2%when the sampled data rate is 30%,which leads to a high image quality.Fourthly,we suggest a data processing procedure for reflective three-dimensional（3D）imaging based on the energy criterion of edge detection for multilayer objects,which are prone to deformation and bend.At first,a correlation operation is performed to align the first peak of echoes,and a time-domain window is applied to acquire the signal from the first layer.Repeat the procedure for the second peak to acquire the signal from the second layer,and so on,until the signals from all layers are separated.The background is estimated by an average method and removed from the signal of each layer.Then the data are transformed into the frequency domain by fast Fourier transform and the image reconstruction is completed using frequency-domain data in specially selected bands.Specifically,we propose the energy-based-edge-detection（EBED）criterion for selecting the frequency bands for the image reconstruction.The experimental results comparing with other selecting criteria have shown that the proposed EBED criterion can provide higher efficiency in the bands selection and leads to higher quality for the constructed image.Next,a total variation method is applied to suppress noise and preserve useful details of the image.At last,the complete reflective 3D THz imaging is obtained by stitching images of all layers.Fifthly,an improved MLEM method with an adaptive penalty function is proposed to deal with the degradation caused by the noise accumulation during the iterative reconstruction of the image.A quadratic function is adopted as the penalty function and the step size is adjusted dynamically with the error between the projection of the reconstructed image and the acquired data,which leads to a higher convergence rate and image quality.Then,this algorithm is examined through a 3D THz imaging to a toy car model,as an example of multi-layer objects.The reconstructed images are further processed using enhancement methods including contrast enhancement,image noise reduction,and morphology filtering,and the 3D terahertz tomography is obtained by stitching images of all layers.At last,a THz computed tomography method based on neural network is proposed.This method utilizes the powerful mapping ability of neural network to map the acquired data into pixel intensities with fewer training samples and outperforms the popular filtered back projection algorithm. |
PDF Full Text Request