Research On Security Inspection Image Processing Method Based On Terahertz Active Imaging

Terahertz(THz)signals are now used in a wide range of fields,including medicine,telecommunications,security monitoring,and quick sensing and imaging.Along with the most evident aspects of the possible applications,there is also the negative end of the spectrum like X-rays which adversely affect biological tissues.On the other hand,Terahertz imaging technology has the advantages of rapid imaging,strong penetration,and harmless to the human body hence widely used in a variety of security environments and has become an alternative technology for X-ray imaging.A well-known problem relating to the THz domain has to do with the scarcity of THz images.Moreover,currently,false detection of objects especially concealed illegal,dangerous,and prohibited objects like weapons and non-weapons in bags,books,wood,etc.as often witnessed in airports,subway stations in terahertz images is the main challenge.Another challenge relates to the low resolution of THz images of which noise is an integral factor.Commencing by resolving the current challenge of scarcity of THz image datasets,400 unprocessed security terahertz images were generated involving instances of the knife,scissors,blade,or screwdriver covered with another material or directly hidden in cases such as wooden box,paper box,a plastic bag or laptop bag.Performing augmentation processes,a dataset comprising 4000 THz images was accumulated for training and for future research work analysis knowing security weapon gadgets are also under regularization by law.In this work,a one-stage object detection model known as YOLOv3 is utilized to resolve the challenge of false detection.This is accomplished by the optimization of the backbone layer at the neck and head sections.Specifically,two additional heads were integrated into the structure.Rigorous experimentation and comparison with existing object detection models show that the new model designed outperforms other existing models.In particular,the modified backbone model CSPDarknet53-SPP and CSPDarknet53-PANet-SPP is 3.1 and 2.5times faster respectively,than that of the YOLOv3 model with Darknet as the backbone as well as an improved accuracy,hence increased in accuracy and speed increases the recognition.Further addressing the challenge,a generative adversarial network(GAN),network-assisted deep learning method is integrated into a newly designed YOLOv3 model for GAN deblurring.Consequently,image quality is improved taking into cognizance peak signal–to-noise ratio(PSNR)and structural similarity(SSIM)values that greatly improved by 42.32% and 53.57%.Apart from the focus on this work towards object detections,true precision,true accuracy,enhanced speed on localization targeted weapon or non-weapon to demystify falsehood information at various security check-points,this work additionally tackles the noise accompanied problem inherent in electronic cameras system of which terahertz camera is not immune to.This work,therefore,advances different denoising schemes under spatial and frequency as well as low and high band-pass filters.Nonlinearity filters such as mean and median filters to denoise terahertz image under the spatial domain yielded a positive outcome after sliding window of 5x5 and 7x7 filters using PSNR and SSIM indicators respectively on salt & pepper and Gaussian noise.At the frequency domain,the wavelet transform of the continuous and discrete wavelet transform filters such as orthogonal and biorthogonal filter families show their superiority in the terahertz image.Denoising and enhancement approach was achieved after testing their signal-noise-ratio(SNR)decomposition and reconstruction level of the coefficient filters using bior4.4 and sym4.0 on salt &pepper and Gaussian noise on the terahertz image.