Research On Super Resolution Algorithm Of Passive Millimeter Wave Images Based On Deep Learning

The millimeter-wave passive detection and imaging system passively receives the electromagnetic waves radiated from the target scene,and uses the radiation differences of different materials for imaging.Because of its all-weather and quasi-all-weather working ability,and its non-radiation,non-contact,high safety and other characteristics,it plays an important role in the civilian and military fields.However,the original passive millimeter-wave image was seriously polluted by noise,lack of detailed information,and low resolution,which could not meet the needs of subsequent target detection and recognition.The super-resolution reconstruction of the image through the algorithm can improve the resolution of the image.However,for the specific field of passive millimeter-wave images,the reconstruction effect of the traditional super-resolution algorithm is not ideal.Therefore,researching a super-resolution algorithm for passive millimeter wave images has important practical significance and application value.In this paper,based on the characteristics of passive millimeter wave images and the characteristics of data sets,an image super-resolution algorithm based on deep learning is studied.The specific work is as follows:(1)In view of the problem that the interpolation method and reconstruction method are not good for the reconstruction of millimeter wave images,this dissertation analyzed the advantages of deep learning algorithms.On the basis of understanding the imaging principle of the millimeter wave passive detection system,the degradation factors in the millimeter wave imaging process were analyzed,and the point spread function of the system was measured.(2)In order to solve the problem of the lack of passive millimeter wave image data set,this paper proposed that according to the imaging process of the passive millimeter wave detection system,using the clear image and the point spread function of the system to interact,the simulated millimeter wave image was obtained to expand the data set.Based on the idea of migration learning,the simulated millimeter wave image was used to pre-train the deep learning network,and then the passive millimeter wave image training set was used to fine-tune the pre-trained network to obtain a network weight suitable for passive millimeter wave image reconstruction.(3)In view of the inherent characteristics of millimeter wave images,such as serious noise pollution,low signal-to-noise ratio,and lack of detailed information,this dissertation proposed a deep convolution residual network.The deconvolution layer was used to replace the traditional interpolation pretreatment operation to reduce the impact of noise on image reconstruction and improve the real-time performance of the network.The method of deepening the network and using small convolution kernel was proposed to increase the sensing field of network,improve the feature extraction ability and nonlinear fitting ability of network.The residual learning network was used to solve the gradient disappearance/explosion problem caused by the deepening of the network,and the sparsity of the network was enhanced.The experimental results verified the effectiveness of the proposed algorithm for super-resolution reconstruction of passive millimeter wave images in different scenes,and verified the superiority of the training method in this paper.