Detection Transformer Framework for Recognition of Heavily Occluded Suspicious Objects

Abstract

Baggage scanning for potential dangers has become a significant worldwide concern in the aviation industry. The manual process of identifying prohibited items can be tedious and chaotic. Researchers have created automated systems to detect potential threats in baggage using X-ray scans, but these systems can still miss items that are hidden or surrounded by clutter. This paper proposes a novel Detection Transformer (DETR) framework for detecting and classifying highly cluttered suspicious items. The proposed framework consists first of extracting the features from the CNN backbone using object proposals that are obtained based on coherent contour maps. These weights are then passed to the CNN model in the DETR to extract the features from the original scan and, therefore, enhance the feature extraction process. In this stage, we feed the transformer encoder-decoder with the representative features for predicting cluttered and concealed prohibited items bounding boxes. The proposed framework has been rigorously evaluated and tested using a total of 47,677 X-ray scans from the publicly available PIDray dataset, where it outperformed the state-of-the-art scheme by 2.10%, 4.05% and 3.64% in terms of mean average precision for easy, hard, and hidden subsets from PIDray dataset, respectively. Keywords: Industries, Neural networks, Virtual environments, Manuals, Detectors, Feature extraction, Transformers