Research On Millimeter Wave Passive Imaging Super Resolution Algorithms Based On Deep Learning

In recent years,the anti-terrorism situation at home and abroad has become increasingly grim,and harmful public incidents have occurred from time to time,which puts forward new and higher requirements for security tasks and new security technologies in the new era of our country.Millimeter wave passive detection imaging technology has excellent technical performance and has important application value in the above fields.However,due to the development cost and the production technology level of some components,the resolution of the imaging results obtained by the system is low,and it can not meet the corresponding processing and operation of detection and recognition.An effective and important way to improve image resolution is imaging super-resolution processing.In view of the specific application background of millimeter wave band imaging,the applicability of the traditional super-resolution algorithm is not good.Therefore,the research and design of new applicable algorithms has important technical significance and application value.Based on the research projects in related fields,this thesis mainly studies the millimeter wave imaging super-resolution algorithm based on deep convolution neural network and based on generative adversarial network.The main contents include:(1)Aiming at the problem that the traditional super-resolution algorithm is difficult to process millimeter wave images,the advantages of the end-to-end super-resolution algorithm based on deep learning are analyzed from the aspects of algorithm complexity and reconstruction effect.The estimation of degradation model of MMW passive detection imaging system is studied.The degradation process of the system and its related influencing factors are analyzed,and the degradation model of the estimation is obtained.(2)Aiming at the problems of shallow model,poor network performance and poor real-time performance of existing deep learning super-resolution algorithms,a deeper convolution neural network super-resolution algorithm for millimeter wave images is proposed through the strategy of continuously deepening network structure.Residual module,recursive module and fast convolution module are designed to solve various problems caused by network deepening and improve real-time performance.The algorithm has better performance,faster speed and can be implanted into the imaging system.(3)Aiming at the smoothing effect brought by the mean square error of the deep learning superresolution algorithm as the network loss function,the perceptive loss function was studied and improved: the loss calculation was carried out using the features before the activation function processing to avoid the adverse effect of feature sparsity on the reconstruction accuracy after activation.(4)Aiming at the problem that the visual effect of the reconstructed result of the network trained with mean square error as loss function is not good under the condition of large super-resolution factor,a super-resolution algorithm based on generative adversarial network for millimeter wave image is proposed.A dense connection residual unit is designed,and the network is trained with improved perceptual loss and network fusion strategy.When the super-resolution factor is large,the super-resolution reconstruction results of millimeter wave images not only have higher evaluation index values,but also have better human visual effect.The simulation and experimental data support the effectiveness and superiority of the proposed algorithms,and verify the feasibility of the depth learning method in millimeter wave image super-resolution processing.