PROHIBITED OBJECT RECOGNITION SIMILARITY-BASED USING SIAMESE NETWORK IN X-RAY BAGGAGE INSPECTION

• Main Author: Kusuma Putra, Rizal
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/73913
• Institution: Institut Teknologi Bandung

Baggage and passengers are two inseparable elements, which is why passenger baggage screening is a crucial process in transportation, especially in air travel, sea vessels, trains, and others. In the baggage screening process, the main source currently used is human professionals. However, human professionals are subject to performance changes caused by fatigue, emotions, and other factors. Therefore, an automated system is needed to assist humans in identifying hazardous objects present in passenger baggage. This research proposes an automated system that assists humans in identifying hazardous objects. The study adopts a different approach from the typical object recognition approach, which is the similarity approach. The similarity approach overcomes the limitations of the commonly used object recognition approaches (classification/probability approaches) in that it can identify partially occluded objects, while classification/probability approaches require complete object samples during the training process to achieve recognition. In this research, three main experimental scenarios were conducted. Firstly, an experimental scenario was conducted to implement Region Proposal Network (RPN) as a detector. Secondly, an experimental scenario was conducted to implement various types of deep learning architectures and Siamese Neural Networks (SNN). Thirdly, an experimental scenario was conducted to combine the RPN and SNN architectures. The study utilized image size of 75x75x3, batch size of 32, and several deep learning encoders such as ResNet50, ResNet101, EfficientNetB7, and InceptionV3. The research results showed that the SNN achieved the best performance with an accuracy score of 88%, while conventional deep learning models only achieved an accuracy score of 74%. This research demonstrates that the SNN model can identify novel objects that have not been seen in the training data, eliminating the need to retrain the model.