A Research On Passive Imaging Super Resolution Algorithm Based On Deep Learning

Passive detection and imaging technology in millimeter wave / terahertz band has important application value in the fields of security inspection,anti-terrorism early warning,battlefield reconnaissance and so on because it can penetrate clothes and camouflage covers,show the physical characteristics of different materials,and do not generate radiation.However,due to the influence and limitation of signal radiation transmission characteristics and device technology level,the imaged images often have the disadvantages of low resolution and noise pollution.Therefore,it is necessary to optimize and process the image through preprocessing technology and super-resolution processing technology.According to the characteristics of millimeter / terahertz imaging images,this paper studies the image preprocessing method of passive imaging and the super-resolution processing algorithm based on depth learning.The main work of this paper is as follows:1.The characteristics of millimeter wave / terahertz imaging and the causes of noise are analyzed and discussed,and the composition of millimeter wave / terahertz detection and imaging system is studied.2.Aiming at the influence of noise on millimeter wave / terahertz wave imaging image,a simple background calibration method is adopted to suppress strip noise.Then,alpha rooting method is used to improve the block matching and three-dimensional filtering denoising algorithm(BM3D).Furthermore,the simple impulse function(sigmoid function)is used to balance the gray value of the image to further enhance the clarity of the image.Through a series of preprocessing methods,the noise of the image is effectively reduced.3.To solve the problem that the number of terahertz image samples is insufficient and it is difficult to complete the training in deep learning,this paper uses the style migration technology to improve the super-resolution algorithm using a generative adversarial network(SRGAN).The improved algorithm inputs the original image and the reference image which is similar to the original image and in the same imaging field into the neural network of the super-resolution processing algorithm,and jointly inputs it into the reconstruction network for image reconstruction after the block replacement step.In the training process,the improved algorithm uses a new loss function to restrict the amplitude of migration.The experiment results show that this method reduces the possible artifacts and noise in the application of deep learning super-resolution algorithm in the field of passive imaging,and improves the application effect of super-resolution algorithm.