Research On Preprocessing And Target Detection Algorithm For Terahertz And Millimeter Wave Passive Imaging

In the 21 st century,under the influence of terrorism,China faces a more serious public security situation.Therefore,it is more important to maintain social stability in this special period.The security checking as an important security barrier in public places such as airports,stations and stadiums,the development of relevant technologies is of great significance.Terahertz and millimeter wave passive detecting imaging technology has the advantages of a good penetrability,no radiation,non-contact,etc,and has a extensive application prospect in the field of security.The image preprocessing and target detection are the core of the technology,which affect the function of the detection system.In this thesis,based on the existing scientific research projects,the algorithm of terahertz passive detection imaging image preprocessing is studied,mainly including image denoising and enhancement.At the same time,based on the detection problem of hidden targets in current millimeter wave images,the research and improvement of traditional target detection methods are explored.According to the characteristics of fast,accurate and low computation in the application process of security check,a fast target detection algorithm based on deep learning is proposed.The main contents are as follows:(1)Based on the blackbody radiation theory,the radiation characteristics of terahertz band and the principle of passive detection and imaging are discussed.The structure and working principle of the terahertz passive detection imaging system are introduced.(2)By analyzing the causes of fringe noise and speckle noise in terahertz image,the preprocessing method of terahertz image is studied.This thesis mainly studies the background calibration denoising method,based on which the wavelet transform and the weighted anisotropic filtering diffusion theory are used to further improve the method and effectively eliminate the noise.At the same time,aiming at the shadow problem in the image,the suppression effect of the classical image enhancement algorithm on the shadow is compared and analyzed through the simulation test.The simulation test results show that the power transformation can effectively suppress the shadow to a certain extent.(3)This thesis studies and analyzes the target detection algorithm based on threshold segmentation,proposes a target detection algorithm based on Twin-Otsu segmentation(Target Detection Algorithm based on Twin-Otsu Segmentation,TDA-TOS)by improving the segmentation algorithm.Aiming at the problem of high false alarm rate in traditional target detection algorithms based on threshold segmentation,the convolutional neural network is proposed to correct the detection results,which can effectively reduce the false alarm rate.(4)In view of the need for fast and accurate detection of hidden dangerous targets in the security field,a deep learning-based target detection algorithm “ID-SSD(Single Shot MultiBox Detector with Improved DenseNet)” is proposed.It adopts SSD algorithm as the basic network model,uses lightweight and efficient feature extraction network,adjusts loss function and a prior boundary box generation strategy,and uses migration learning theory to solve the training problem of small sample set.The simulation results show that the algorithm is fast,accurate and low computation.