Research On Sample Feature Recognition Algorithm Based On Terahertz Time Domain Spectroscopy

The energy level of vibration and rotation for biological and chemical macromolecules is most among terahertz(THz)waveband.Meanwhile,THz wave radiation possesses unique advantage to penetrate vast majority nonmetallic nonpolar materials.Therefore,there is an extensive application of spectral measurement and object recognition in biomedicine,public safety,nondestructive examination,quality inspection,chemical correlation and environmental monitoring fields by using terahertz time domain spectroscopy(THz-TDS)technology.In fact,THz spectrum possesses typical hyper-spectral data feature with high spectral dimension and large amount of data,thus how to survey and mining useful information from massive THz spectrum data is a great challenge in THz-TDS technology application and development.Based on the THz spectrum theory and sample spectrum data obtained by THz-TDS detection technology,this thesis carries out the research according to two key issues in spectrum recognition.One is the recognition of material with similar spectral characteristic.To improve the target recognition accuracy by THz spectrum,we extracting the main differences from full spectrum by data-driven to investigate new spectrum recognition method.The other is the target characteristic recognition fail problem due to spectrum image noisy and low optical resolution.Based on the intensive study of THz imaging mechanism and fully develop the useful information from the image and spectrum integration feature of THz spectrum image,we taking advantage of target differential spectrum clustering,image denoising and high-resolution image similar matching to lay foundations for image target super-resolution recognition.Main research work and innovations are as follows:? We propose a parameter optimization model based on support vector machine(SVM)and adaptive particle swarm optimization(APSO)to resolve the material recognition problem with similar spectrum features.Indeed,the THz spectrum scanning and sampling procedure is quite complex and easy to be affected by dynamic environment,the optimal parameter value of classifier is also dynamic changing when execute the classification training for complex diversity materials,which may result in a trapping in local optimal or endless loop for PSO algorithm to optimize the classifier parameter.Therefore,this thesis improves the traditional PSO algorithm and designs a new adaptive PSO method(APSO)by introducing perception particle,so that to supervise and update the parameter optimization.First of all,this mechanism randomly selects several particles from the PSO feasible solution space as the perception particle.Then,during the PSO iterative optimization process,compute the sum of adjacent iteration fitness difference for all the perception particles to describe the severity degree of environment changing.Thirdly,setting up a response threshold,and trigger the threshold update response when the sum of particle fitness difference value is bigger than the response threshold,so as to force the parameter optimization jump out of current local optimal or endless loop.Comparative experiments of several SVM models on genetically modified cotton seeds sample illustrates that our new algorithms can improve the model recognition accuracy and the adaption of dynamic environment,reduce the parameter optimization iteration times and obtain a higher efficiency.? In terms of the spectrum recognition,this thesis put forward a spectrum clustering recognition model based on t-distribution stochastic neighborhood embedding(t-SNE)to resolve the problem that THz spectrum image resolution is too low that it can't be effectively used for sample defect feature recognition.The model realizes the reduced dimension clustering of difference spectrum drawn from the imaging spectrum data sets,so as to judge the sample defect feature if or not among the low dimension space.Mapping the spectrum data samples from high dimension space to low dimension space by the manifold learning algorithm t-SNE,and measuring the distance invariant similarity by conditional probability distribution,so as to achieve a visible observation of sample features in low dimension space.Comparative experiments on image indicate that our model can judge there is different sample defect feature or not through spectrum clustering,realizing the pre-detection for image analysis.? This thesis proposes a new image denoising model based on ant colony optimization(ACO)edge detection and compressive sensing denoising(CS)to resolve the problem that traditional digital image denoising methods are prone to blur some vital edge information of THz spectrum image,and will affect the accuracy of image target recognition.The basic idea of the new model is as follows:firstly,the spectrum image is divided into edge and non-edge images by ACO,and only the non-edge image is denoised to protect the important edge features of the image.Meanwhile,based on the CS theory,the sparse representation of non-edge image is obtained by orthogonal transformation,then the local fast Fourier transform image reconstruction algorithm is used for image denoising.Compared to typical image denoising models,our proposed model contains best effect and more utility.? This thesis proposes a super-resolution image matching model based on image SIFT feature extraction and K-means clustering to retrieve a group of higher quality source image from image data sets.The new model we proposed solves the problem of limited improvement of image quality due to the randomness of image source selection when image fusion and super-resolution reconstruction are used to process degraded spectral images.To be specific,the Fourier transform method based on Blackman window frequency domain interpolation is utilized to improve the frequency domain spectrum resolution.Then the THz spectrum image library is built by multiple parameter information imaging of the frequency and time domain.Furthermore,a high-quality image is chosen as matching image according to the sample spectrum cut-off frequency and image objective evaluation comprehensive index analysis.Finally,a group of high-resolution source images are retrieved according to the SIFT-Kmeans model and vector cosine similarity measurement computation.Experiment on multiple samples illustrate that our proposed image quantitative analysis system is guarantee the source image quality and prove the image super-resolution reconstruction is quite relevant.