Detection of concealed electronic circuits in X-ray images using deep learning methods

This thesis focuses on the detection of explosive circuits hidden in electronic devices, such as laptops, using X-ray imaging technologies, which are commonly used for security purposes in public institutions. The detection of prohibited items by security personnel through X-ray images may not be efficient due to issues such as time constraints, labor, and lack of expertise. Therefore, various methods have been developed to automate this process and eliminate human factors. Deep learning methods hold the potential to overcome these issues by automating image analysis and object recognition. In the first phase of the study, an original dataset was created for detecting explosive circuits, and classification was performed using deep learning models on this dataset. However, due to the small size and high complexity of the dataset, various problems arose during the testing phase. To address these issues, in the second phase, a method was developed to improve classification performance by combining feature maps from deep learning models. In particular, 11 different models based on Convolutional Neural Networks (CNN) were designed for classification purposes. The first challenge in the study was the small size of the dataset, which led to overfitting during testing, resulting in lower test performance compared to training performance. To mitigate this issue and improve performance, a classifier method that requires shorter training time and is resistant to overfitting was proposed. Thus, a feature fusion method with a Random Weight Network (RWN)-based classifier was introduced, and its performance was compared with the deep learning methods in the literature, as well as with a method where only the classifier was changed, for explosive detection. In the final section of the study, an analysis was conducted using RWN and feature fusion. This analysis aimed to evaluate the performance and effectiveness of each approach under different intermediate layer and hyperparameter values. As a result, the contribution of the developed feature fusion method to success was demonstrated through experimental results.