Object Detection in X-ray Images Using Transfer Learning with Data Augmentation
| Main Authors: | Galvez, Reagan L.; Gokongwei College of Engineering, De La Salle University, Dadios, Elmer P.; Gokongwei College of Engineering, De La Salle University, Bandala, Argel A.; Gokongwei College of Engineering, De La Salle University, Vicerra, Ryan Rhay P.; Gokongwei College of Engineering, De La Salle University |
|---|---|
| Format: | Article info application/pdf eJournal |
| Bahasa: | eng |
| Terbitan: |
International Journal on Advanced Science, Engineering and Information Technology
, 2019
|
| Subjects: | |
| Online Access: |
http://insightsociety.org/ojaseit/index.php/ijaseit/article/view/9960 http://insightsociety.org/ojaseit/index.php/ijaseit/article/view/9960/pdf_1286 |
Daftar Isi:
- Object detection in X-ray images is an interesting problem in the field of machine vision. The reason is that images from an X-ray machine are usually obstructed with other objects and to itself, therefore object classification and localization is a challenging task. Furthermore, obtaining X-ray data is difficult due to an insufficient dataset available compared with photographic images from a digital camera. It is vital to easily detect objects in an X-ray image because it can be used as decision support in the detection of threat items such as improvised explosive devices (IED’s) in airports, train stations, and public places. Detection of IED components accurately requires an expert and can be achieved through extensive training. Also, manual inspection is tedious, and the probability of missed detection increases due to several pieces of baggage are scanned in a short period of time. As a solution, this paper used different object detection techniques (Faster R-CNN, SSD, R-FCN) and feature extractors (ResNet, MobileNet, Inception, Inception-ResNet) based on convolutional neural networks (CNN) in a novel IEDXray dataset in the detection of IED components. The IEDXray dataset is an X-ray image of IED replicas without the explosive material. Transfer learning with data augmentation was performed due to limited X-ray data available to train the whole network from scratch. Evaluation results showed that individual detection achieved 99.08% average precision (AP) in mortar detection and 77.29% mAP in three IED components.