Research On Super-resolution Reconstruction And Object Detection Based On Millimeter Wave And Terahertz Image

In our country,public safety has always been one of the most important issues for the public and the country.There are some security inspections in our transportation hubs.However,the current security inspections of the human body still rely on the use of inefficient metal detectors and other methods.Millimeter-wave and terahertz has high permeability to fabrics and high reflectivity to metals and other substances.It can realize non-contact detection of dangerous goods carried by the human body and in luggage.Then,the millimeter-wave and terahertz can be detected by the target detection algorithm.The image detection can realize automatic recognition.Because of the limitations of equipment performance and imaging environment in practical applications,millimeter-wave and terahertz images have the characteristics of low resolution and blurred edges.This paper studies the super-resolution reconstruction of millimeter-wave and terahertz images to improve image quality.At the same time,because the number of parameters of the target detection model based on deep learning is too big,which is not conducive to deployment,this paper studies the lightweight transformation of the target detection algorithm.Aiming at the characteristics of millimeter-wave and terahertz images with low resolution and blurred edges,this paper proposes an edge-enhanced super-resolution method based on SRGAN,which improves SRGAN in terms of network structure and loss function.In the network structure part,the influence of network depth on millimeter wave image and terahertz image reconstruction is analyzed,the number of network layers of the generator network is reduced,and a dense connection structure is added to retain the shallow characteristics of the previous network layer.In the loss function part,the edge features of the image are extracted through the Laplace operator,and the edge loss between the reconstructed image and the original image is calculated based on this,so that the network pays more attention to the edge information of the image.In order to realize real-time security inspection,this paper first verifies the detection effect of YOLOv4 algorithm and YOLOv4-Tiny algorithm on millimeter-wave terahertz images,and then performs lightweight transformation on the basis of YOLOv4 algorithm.The YOLOv4 algorithm has high detection accuracy,but the detection speed is slow,while the accuracy of YOLOv4-Tiny is severely reduced due to insufficient feature fusion.Based on the feature fusion module of YOLOv4,this paper replaces the feature extraction module in YOLOv4 with the feature extraction module in Mobile Net to construct a YOLOv4-LW network,and replaces the standard convolution in the feature fusion module with deep separable convolution,which further reduces the amount of parameters.After experimental verification on millimeter-wave terahertz human inspection and physical inspection datasets,the super-resolution reconstruction method proposed in this paper can effectively realize the reconstruction of millimeter-wave and terahertz images and achieve better results than the original SRGAN.The YOLOv4-LW network proposed in this paper can achieve a balance of detection speed and accuracy,which is faster than YOLOv4,more accurate than YOLOv4-Tiny,and meets the needs of real-time security inspection.