DINOray: A Tightly Integrated De-occlusion Model for the Detection of Prohibited Items

In today’s society, the increasing occurrence of terrorist actsand other criminal behaviors highlights the crucial necessity of securityscreening to safeguard areas from potential threats. However, most of theexisting mainstream target detection methods are designed for process-ing natural images, and do not take into account the special differencethat X-ray security images have perspective. The prohibited items withinthese security images exhibit size and shape variations under variousimaging angles, often rendering them uninformative to the models. Inaddition, the security images frequently display varying degrees of over-lapping and occlusion between targets and backgrounds, which makes theaccurate detection of prohibited items more difficult. To address thesechallenges in security screening tasks, this paper proposes a novel modelbased on DINO, which is referred to as DINOray. It integrates a Cross-Scale Tight Integration module and an Occlusion Removal module. Bymitigating noise and occlusion, DINOray extracts more salient informa-tion about prohibited items and tightly integrates it with the output ofthe encoder, which includes global features, in order to supplement lo-cal features. These enhanced features are then passed to the decoder toenhance the detection accuracy of the model. The experimental resultsshow that DINOray outperforms DINO by achieving 3.5% and 1.1% en-hancements in the detection accuracy of prohibited items on the PIDray and CLCXray open-source security screening datasets, respectively. Thisdemonstrates its superior performance in the security screening field.