Abstract

Visual inspection of dual-energy X-ray radiographic images of cabin baggage requires high performance, but is hindered by various challenges such as low target prevalence, variability in target visibility, possible presence of multiple targets, and security personnel fatigue and inattention. Artificial intelligence (AI) techniques, particularly deep convolutional neural networks (CNNs), have shown promise in improving the automatic detection of explosives, even with low-resolution radiographic images, especially in high baggage throughput scenarios. In this paper, we focus on the detection of detonators as components of improvised explosive devices. The proposed approach involves comparing two experiments implemented in a deep CNN architecture using TensorFlow and Keras libraries. In the first experiment, raw dual-energy radiographic images without any enhancement were used. The second experiment includes three methods for contrast enhancement and feature extraction: the Contrast Limited Adaptive Histogram Equalization (CLAHE) method, the wavelet transform-based method, and the mixed CLAHE RGB-Wavelet method. In the latter two methods, Haar, Db2, Coif2, and Sym2 mother wavelet functions at two levels (HH and HL) were employed. The analysis of results focuses on a comparative study of performance measures such as accuracy, precision, recall, and F1 score. It was found that the preprocessing methods used in experiment 2, for the two evaluated classes (detonator and no detonator), achieved higher accuracy compared to the raw radiographic images used in experiment 1 (98.08%). The highest accuracies in experiment 2, with a value of 100%, were obtained with the CLAHE method (green channel in grayscale, blue channel in grayscale, and RGB channels) and the wavelet transform method with Haar mother wavelet at two levels HL