7 Mar 2024 | Mingyuan Li, Tong Jia, Hao Wang, Bowen Ma, Shuyang Lin, Da Cai, and Dongyue Chen
The paper "AO-DETR: Anti-Overlapping DETR for X-Ray Prohibited Items Detection" addresses the challenge of detecting prohibited items in X-ray images, which often exhibit significant overlapping and edge blurring. The authors propose a novel model called Anti-Overlapping DETR (AO-DETR), which is based on the state-of-the-art DETR-like model DINO. To enhance the model's ability to handle overlapping phenomena, the authors introduce two key strategies: Category-Specific One-to-One Assignment (CSA) and Look Forward Densely (LFD).
1. **Category-Specific One-to-One Assignment (CSA)**: This strategy constraints category-specific object queries to predict only the features of specific prohibited items, enhancing the model's ability to extract relevant features from overlapping foreground and background. It ensures that queries are specialized for specific categories, improving the accuracy of feature extraction.
2. **Look Forward Densely (LFD)**: This scheme improves the localization accuracy of reference boxes by densely transmitting gradients through multiple decoder layers. It helps in refining the localization of blurry edges, particularly in mid-to-high-level decoder layers, leading to more precise edge localization.
The authors conduct extensive experiments on the PIXray and OPIXray datasets, demonstrating that their proposed methods significantly outperform state-of-the-art object detectors. The results show that AO-DETR achieves higher mean average precision (mAP) and better recall rates, especially in scenarios with severe overlapping and edge blurring.
The paper also includes a detailed analysis of the impact of CSA and LFD on the model's training process, using metrics such as foreground instability score (FIS) and Hungarian matching instability (IS). The visualization analysis further illustrates how category-specific object queries focus on regions most similar to the features of their responsible categories, enhancing the model's ability to handle overlapping scenarios.
Overall, the paper contributes to the field of X-ray prohibited item detection by providing a robust and effective solution to the challenges posed by overlapping and edge blurring in X-ray images.The paper "AO-DETR: Anti-Overlapping DETR for X-Ray Prohibited Items Detection" addresses the challenge of detecting prohibited items in X-ray images, which often exhibit significant overlapping and edge blurring. The authors propose a novel model called Anti-Overlapping DETR (AO-DETR), which is based on the state-of-the-art DETR-like model DINO. To enhance the model's ability to handle overlapping phenomena, the authors introduce two key strategies: Category-Specific One-to-One Assignment (CSA) and Look Forward Densely (LFD).
1. **Category-Specific One-to-One Assignment (CSA)**: This strategy constraints category-specific object queries to predict only the features of specific prohibited items, enhancing the model's ability to extract relevant features from overlapping foreground and background. It ensures that queries are specialized for specific categories, improving the accuracy of feature extraction.
2. **Look Forward Densely (LFD)**: This scheme improves the localization accuracy of reference boxes by densely transmitting gradients through multiple decoder layers. It helps in refining the localization of blurry edges, particularly in mid-to-high-level decoder layers, leading to more precise edge localization.
The authors conduct extensive experiments on the PIXray and OPIXray datasets, demonstrating that their proposed methods significantly outperform state-of-the-art object detectors. The results show that AO-DETR achieves higher mean average precision (mAP) and better recall rates, especially in scenarios with severe overlapping and edge blurring.
The paper also includes a detailed analysis of the impact of CSA and LFD on the model's training process, using metrics such as foreground instability score (FIS) and Hungarian matching instability (IS). The visualization analysis further illustrates how category-specific object queries focus on regions most similar to the features of their responsible categories, enhancing the model's ability to handle overlapping scenarios.
Overall, the paper contributes to the field of X-ray prohibited item detection by providing a robust and effective solution to the challenges posed by overlapping and edge blurring in X-ray images.