PODX: A Comprehensive Framework for Detecting Prohibited Items in Security X-Rays

• Published in: 2024 9th International Conference on Communication and Electronics Systems (ICCES) pp. 946 - 951
• Main Authors: Shivane, Ashwini S., Bhoite, Sachin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.12.2024
• Subjects: Accuracy; CNN; Convolutional neural networks; Decision Trees; Feature extraction; feature selection; Gradient Boosting; LASSO; Machine learning algorithms; Nearest neighbor methods; Random forests; Reliability; Security; Support vector machines; X-ray imaging
• DOI: 10.1109/ICCES63552.2024.10859353

The increasing need for accurate and reliable detection of prohibited items such as guns, knives, and pliers in X-ray security images underscores the importance of advanced frameworks to enhance security screening processes. This study introduces Prohibited Object Detection in X-rays (PODX), a comprehensive two-stage framework aimed at improving detection capabilities through advanced feature extraction and classification techniques. The framework employs a two-stage approach: in the first stage, three convolutional neural network (CNN) architectures-ResNet, GoogLeNet, and DenseNet-are utilized to extract features, which are further refined using LASSO (Least Absolute Shrinkage and Selection Operator) regression to select the most significant features for classification. In the second stage, multiple machine learning algorithms, including Random Forests, Gradient Boosting, Support Vector Machines (SVM), and K-Nearest Neighbors (KNN), are implemented to classify prohibited items based on the refined features. The performance of these algorithms is evaluated using metrics such as train accuracy and test accuracy, providing a comprehensive assessment of their strengths and weaknesses. The PODX framework not only enhances the accuracy and reliability of detecting prohibited items in security applications but also offers valuable insights into the comparative effectiveness of various classification techniques in X-ray image analysis.