AO-DETR: Anti-Overlapping DETR for X-Ray Prohibited Items Detection Prohibited item detection in X-ray images is a critical task in security inspections. Due to overlapping phenomena in X-ray images, existing detectors face challenges in accurately detecting prohibited items. To address this, the authors propose AO-DETR, an anti-overlapping DETR model based on DINO, a state-of-the-art general object detector. AO-DETR introduces two key strategies: Category-Specific One-to-One Assignment (CSA) and Look Forward Densely (LFD). CSA enhances the ability of category-specific object queries to extract features from overlapping foreground and background by assigning them to specific category labels. LFD improves the localization accuracy of reference boxes by densely transmitting gradients to multiple decoder layers, enabling more accurate edge localization. Extensive experiments on the PIXray and OPIXray datasets show that AO-DETR outperforms state-of-the-art object detectors, demonstrating its effectiveness in prohibited item detection. The model achieves high detection accuracy, with AP values of 65.6% on PIXray and 73.9% on OPIXray. AO-DETR also maintains high inference speed, with a frame rate of 54 fps when using a lightweight version based on ResNet-50. The model's performance is further validated through ablation studies, which show that both CSA and LFD contribute to improved detection accuracy. Additionally, the model's stability is evaluated using foreground instability scores (FIS) and object assignment instability scores (IS), which are consistently lower in AO-DETR compared to baseline models. AO-DETR's effectiveness is further demonstrated through visualization analysis, which shows that category-specific object queries focus on regions in images that exhibit the highest similarity to the features of prohibited items of their responsible category. The model's ability to accurately detect prohibited items in overlapping scenarios is validated through comparisons with other state-of-the-art detectors, including general object detectors and prohibited item detectors. The results show that AO-DETR achieves superior performance in terms of detection accuracy and inference speed, making it a promising solution for X-ray prohibited item detection.AO-DETR: Anti-Overlapping DETR for X-Ray Prohibited Items Detection Prohibited item detection in X-ray images is a critical task in security inspections. Due to overlapping phenomena in X-ray images, existing detectors face challenges in accurately detecting prohibited items. To address this, the authors propose AO-DETR, an anti-overlapping DETR model based on DINO, a state-of-the-art general object detector. AO-DETR introduces two key strategies: Category-Specific One-to-One Assignment (CSA) and Look Forward Densely (LFD). CSA enhances the ability of category-specific object queries to extract features from overlapping foreground and background by assigning them to specific category labels. LFD improves the localization accuracy of reference boxes by densely transmitting gradients to multiple decoder layers, enabling more accurate edge localization. Extensive experiments on the PIXray and OPIXray datasets show that AO-DETR outperforms state-of-the-art object detectors, demonstrating its effectiveness in prohibited item detection. The model achieves high detection accuracy, with AP values of 65.6% on PIXray and 73.9% on OPIXray. AO-DETR also maintains high inference speed, with a frame rate of 54 fps when using a lightweight version based on ResNet-50. The model's performance is further validated through ablation studies, which show that both CSA and LFD contribute to improved detection accuracy. Additionally, the model's stability is evaluated using foreground instability scores (FIS) and object assignment instability scores (IS), which are consistently lower in AO-DETR compared to baseline models. AO-DETR's effectiveness is further demonstrated through visualization analysis, which shows that category-specific object queries focus on regions in images that exhibit the highest similarity to the features of prohibited items of their responsible category. The model's ability to accurately detect prohibited items in overlapping scenarios is validated through comparisons with other state-of-the-art detectors, including general object detectors and prohibited item detectors. The results show that AO-DETR achieves superior performance in terms of detection accuracy and inference speed, making it a promising solution for X-ray prohibited item detection.