Abstract:
Baggage inspection for threat objects using X-ray images is a priority task
that is in charge of making the risk of crime and terrorist attacks more reducible.
Nowadays, the checking of baggage is based on a semi-automated system that consists
of both human and also image detection. The main purpose of this thesis is to make the
automatisation system more reliable. This task is mainly dependent on object
detection models and algorithms. However, another part of this problem is mainly
prone to the lack of the data. Furthermore, obtaining an X-ray image dataset with many
types of threat objects is quite difficult. This is why this thesis is composed of two
main parts: data stimulation and object detecting approaches. On the first part, due to
the lack of data, an older dataset containing only four classes of threat objects are used
as a base for the new objects to be stimulated into. In total, the newly simulated dataset
contains seven types of threat objects consisting of: handguns, razor-blades, knife,
shuriken, battery, wires and mortar. After generating the new data, they are processed
and augmented by applying random types of rotations, flippings, zoomings etc. Once
the images are processed, they are passed into transfer learning. Transfer learning
consists of using predefined models for training. The models that are taken into
consideration are: Single Shot Detector, Regions Fully Convolutional Network (R-
FCN) and Faster R-CNN. These models are used by applying different techniques of
feature extraction, such as: Inception-v2, MobileNet-v2 and ResNet101. Combining
the object detection models and object detection architectures in total the images are
trained and tested in five different approaches. In conclusion, the best detection was achieved by the combination of Faster-RCNN detection model and ResNet101 feature
extractor by 87.58% ± 0.75 accuracy.
