Research Of Small Object Detection In Passive Terahertz Human Body Image

With the development of passive terahertz imaging technology,passive terahertz imaging equipment for human security has been widely used in public transport facilities for human security.However,passive terahertz devices are prone to generate random noises during imaging process.The characteristics of objects may be lost during imaging,and passive terahertz images are low contrast and edge blur.All of these have brought great challenge to the object detection in the passive terahertz image for human body security check.It is also necessary to detect the small contraband in the human body security check quickly and accurately in some specific security check scene(such as airport).The small object detection has always been a difficult problem in the detection field.Currently,According to the characteristics of random noise and low contrast in passive terahertz image,the research of various preprocessing algorithms is only to improve the quality of passive terahertz image.The research of feature loss is also less during passive terahertz imaging.In the small object detection,traditional detection methods and simple deep learning detection methods cannot meet the needs of accuracy and speed for small object detection in human body security check.In this paper,aiming at the characteristics of passive terahertz human body security image and the low accuracy of small object detection,we propose an deep learning detection method combining with pre-processing and image fusion technology.The main works are as follows:(1)Because there are random noise,low contrast and edge blur in passive terahertz image,we analyze and compare the effects of various denoising and enhancement algorithms.The experiments show that the median filter combined with the logarithmic nonlinear transformation can effectively suppress the random noise and enhance the contrast of the image,which provides favorable conditions for small object detection.(2)To solve the feature missing problem caused by the hardware of passive terahertz sensor,we analyzes and compares various image registration and fusion algorithms in this paper.The experiments show that the image registration algorithm based on features combined with pixel level image fusion method perform well.(3)To improve the accuracy of small object detection,we propose an optimization SSD network structure(OSSD model)in this paper.In this model,we improve the accuracy of small object detection by optimizing the architecture and default boxes and data augmentation.(4)In this paper,we propose optimization SSD model combining with pre-processing and image fusion technology.The model not only improves the accuracy of small object detection furtherly,but also improves the accuracy of general object detection.The experiment show that the OSSD model combined with preprocessing and image fusion can meet the needs of security check scene.The model also improve the accuracy of small object detection and general object detection,which reaches 75.1% and 77.5% separately.