Full Article: PDF
Scientific Object Identifier: http://s-o-i.org/1.1/TAS-03-95-39
DOI: https://dx.doi.org/10.15863/TAS.2021.03.95.39
Language: English
Citation: Danso, S. A., et al. (2021). Optimizing YOLOv3 detection model using terahertz active security scanned low-resolution images. ISJ Theoretical & Applied Science, 03 (95), 235-253. Soi: http://s-o-i.org/1.1/TAS-03-95-39 Doi: https://dx.doi.org/10.15863/TAS.2021.03.95.39 |
Pages: 235-253
Published: 30.03.2021
Abstract: Terahertz technology is nonionizing radiation consequently posing less human risk. However, its spectroscopy-scanned images are characterized by low-resolution images thereby posing significant challenges when object detection is to be performed in such images. Recently, deep learning-based detection has shown much prospects owing to their highly based computer vision approach for its superior efficiency and easy network parameter optimization. In this paper, we perform a comprehensive analysis of prominent object detection models based on terahertz images regarding concealed dangerous and prohibited objects in bags, books, wood etc. as often witnessed in airports, subway stations etc. By way of boosting the performance coupled with detection accuracy of the models, we expand our initial terahertz images via image augmentation. Experimental results reveal that one-way detection method for hidden weapons and non-weapons is far better than two-way detection methods. Moreover, we achieved a 2% increased accuracy and an increased rate of 2.5 due to the optimization fromYOLOv3.
Key words: Terahertz image, object detection, deep learning, hidden weapon.
|
