Research On Terahertz Security Inspection Image Recognition And Detection Based On Improved Faster R-CNN Network

Between the millimeter wave band and the far infrared band in the electromagnetic spectrum,there is an electromagnetic wave that overlaps the millimeter wave in the low frequency region and the infrared wave in the high frequency region,that is,the terahertz wave,which is a kind of interspersed in electronics and photonics.Between technologies.Terahertz waves have many special properties,such as pulse width spectrum,low energy of photons,strong penetration characteristics for non-polar materials,good coherence,and instantaneous advantages.With the advancement of science and technology,the front-end technology of the core discipline of terahertz has been developed very rapidly,and the terahertz imaging system has gradually been applied to public safety inspection bays such as subway stations,train stations,and bus stations in various cities.Human body imaging detection.Human body images based on visible light cannot detect objects hidden inside human clothing,and terahertz imaging systems have the advantage of being able to penetrate materials such as human clothing and present hidden objects(including metallic and non-metallic materials,etc.).However,the terahertz image generated by the terahertz imaging system has the characteristics of single hue,resolution,sharpness,contrast difference,single sample,and small amount of data.The traditional terahertz image detection method mainly combines the priori knowledge with the classifier For target detection,there is a problem of low detection efficiency and accuracy.At present,the method of detecting objects of interest based on deep learning has a very good detection effect,but requires a large number of data sets as a premise.Therefore,in this paper,in view of the above problems in the terahertz security image,the following three aspects of work are carried out:(1)This paper uses the method of DCGAN to generate adversarial networks to generate terahertz security inspection images to expand the terahertz security inspection image data.For the problem of high cost,the generation of adversarial networks through DCGAN can generate a large number of diverse terahertz security inspection image data,effectively expanding the data volume of training and detection samples,and ensuring that the models obtained after deep neural network training are stronger Robustness.(2)The expanded terahertz security inspection image data and the original terahertz security inspection image data are sent to the ESRGAN enhanced super-resolution generation confrontation network to perform super-resolution reconstruction of the terahertz security inspection image to realize the conversion of low-resolution terahertz security inspection images into High-resolution images to obtain higher visual quality and more realistic and natural texture features,which solves the problems of low resolution and clarity of the original terahertz security image.After the ESRGAN network super-resolves the terahertz security image,the background noise in the terahertz security image still exists,and the contrast between the region of interest and other backgrounds is not obvious enough.This paper further performs threshold processing and linear transformation on the super-reconstructed terahertz security image,filtering the background noise and improving the discrimination between different targets,so that the quality of the processed terahertz security image has been greatly improved The distinction between different goals has been further improved.(3)The classic target detection network algorithm Faster R-CNN network is used as the main structure of the suspicious target in the terahertz security image after the detection process,considering that there are a large number of stacked between suspicious target objects to be detected in the terahertz security image The problem,as well as the defects of the NMS non-maximum suppression algorithm in the original Faster R-CNN network framework when screening the required detection frame(when the two target detection objects to be detected have a high degree of overlap,they will be deleted.Another target frame with low lead to the problem of missed detection),improved NMS algorithm.Introducing the Sigmoid weighting function reduces the confidence of another detection frame that overlaps with the target,instead of directly deleting it,retaining other target detection frames with lower scores,solving the problem of missed detection due to target overlap,and improving the detection of terahertz security The detection accuracy of suspicious objects in the image.