...:::Dr. Mustafa Servet Kıran:::...

Subject and Importance of the Project

The increase in terrorist attacks has made physical security scanning a necessity in private and state properties like airports, shopping centers, and ministries. X-ray imaging technologies have been used to scan people and their belongings, but the process is not efficient and safe due to time constraints, workload, and lack of expert personnel. Detection of dangerous substances and explosive materials is difficult, especially in bags and luggage where items overlap. Expert personnel must follow X-ray images and determine if there is a dangerous substance, which is a challenging task. Therefore, the study proposes an automated process using image processing and deep learning algorithms to detect dangerous substances and explosive materials.

The main purpose of this study is to improve the reliability and effectiveness of security inspections at airports, where human and passenger safety is most essential. In addition, it is aimed to automate the related process in a way to include the detection of hidden explosive circuits in laptop computers and to reduce the dependence on humans. In this direction, the objectives of the research are as follows:

Collecting of a data set consisting of laptop computers and hidden explosive circuits.
Analysis of training and test performance of different deep learning architectures on the collected dataset.
Development an automated system to detect the bomb set-up, that does not requires an expert personnel.

Methods

The proposed system for automated physical security scanning uses x-ray images that are collected from x-ray devices and prepared for image analysis. The images are labeled as normal and abnormal, and a deep learning model is trained to classify new images from the x-ray device. The general framework of the system is shown in Figure 1.

ecd

Fig. 1. The general framework of the proposed system

In this research, foreign electronic circuits were hidden inside laptop computers, and x-ray images of these computers were taken from x-ray devices. The obtained images were used to train deep learning models at different angles and placements. The deep learning models used in this study are InceptionV3, EfficientNet, ResNet18, ResNet50, ResNet101, DarkNet53, DarkNet19, MobileNetV2 and ShuffleNet.

Dataset

A dataset consisting of 6200 X-ray images was created in the training hall of Konya Airport over a period of average 5 months with technical assistance from local official institutions. The dataset included 60 laptops with various electronic circuits such as Arduino uno, nano and Bluetooth circuits concealed in different areas inside the laptop. The images of these devices were taken from different angles and circuit combinations using plastic boxes used in airports, and their label and location information were stored.

x-ray images

Fig. 2. X-Ray images with/without electronic circuits

As seen from the laptop X-ray images with/without electronic circuit in Figure 2, the detection of the hidden circuit is a challenging process.

Results and Discussion

The X-ray images were pre-processed by removing unnecessary parts and the dataset was balanced by considering the ratio of normal and abnormal images. The dataset contained 5094 images, and it was divided into two sections. The first section was used to train the models using the 10-fold cross-validation technique. The second section was used exclusively for testing the models. The experiments were conducted using various deep learning models such as InceptionV3, EfficientNet, ResNet18, ResNet50, ResNet101, DarkNet53, DarkNet19, MobileNetV2 and ShuffleNet. The architecture of the algorithms was not modified, and the results were averaged across all folds.

training process

Fig. 3. Models during the training process

The provided accuracy values in Figure 3 show that different deep learning models were trained on a dataset of X-ray images. All models started with low accuracy but gradually increased during training. The scores for all models were generally high, indicating good performance. EfficientNet and ShuffleNet models achieved the highest accuracy scores in the training phase, while other models such as InceptionV3, ResNet50, DarkNet53, DarkNet19, MobileNetv2, ResNet18 and ResNet101 had lower accuracy scores.

The learning speed of a deep learning model depends on various factors such as model complexity, dataset size, optimizer, loss function, and regularization techniques. The ShuffleNet model appears to have the highest accuracy and reaches it faster than other models, while the other models take more epochs to reach similar levels of accuracy. However, high accuracy on the training data doesn't guarantee good performance on new, unseen data. It's important to also evaluate the model's performance on a validation set to check for overfitting.

validation process

Fig. 4. Models during the validation process

The Figure 4 shows the validation accuracies of different deep learning models on a validation dataset. The ShuffleNet model has the highest peak validation accuracy of 83% and shows consistently high accuracy throughout the range of values provided, suggesting that it is performing well and generalizing well. The InceptionV3 model has the second-highest peak validation accuracy at 80%. The other models, including EfficientNet, ResNet50, DarkNet53, DarkNet19, MobileNetv2, ResNet18, and ResNet101, have peak validation accuracies between 60-65%.

Conclusions and Future works

The study examined the performance of different deep learning models for detecting explosive circuits from X-ray images. A new dataset consisting of X-ray images of laptops with/without an electronic card has been created and 9 deep learning models were applied to classify the laptops as normal or abnormal. The ShuffleNet model performed better than the other models on all comparison metrics. The study's findings can be useful in developing automated systems for security purposes, which can reduce the workload of security professionals and improve accuracy. The deep learning algorithms are applied as proposed in the base study, and hyper parameter optimization for this dataset should be required to improve the classification performance of the models. Moreover, the region-based deep learning models can be applied to the dataset, and these hyper parameter optimization and region-based models will be applied to the dataset in near future.