Core Concepts
This paper systematically evaluates the capabilities of Vision Transformers and hybrid architectures for the task of illicit item detection in X-ray images, demonstrating the remarkable accuracy of the DINO Transformer detector in the low-data regime, the impressive real-time performance of YOLOv8, and the effectiveness of the NextViT backbone.
Abstract
The paper focuses on the task of automated illicit object detection in X-ray images, which is crucial for public safety at critical terminals like airports, train stations, and ports. It evaluates various deep neural architectures, including Transformer-based and hybrid models, on this task.
The key highlights are:
The paper compares the performance of SWIN Transformer, NextViT (a hybrid CNN-Transformer model), and YOLO-based detectors on the SIXray and CFray datasets.
The results show that the DINO Transformer detector achieves remarkable accuracy in the low-data regime of the CFray dataset.
YOLOv8 combined with the NextViT backbone demonstrates impressive real-time inference speed, while maintaining competitive accuracy.
The hybrid NextViT backbone proves effective, outperforming the pure Transformer SWIN in terms of the stricter mAP@50-95 metric on the SIXray dataset.
The paper suggests that end-to-end Transformer architectures may face challenges in the highly specific X-ray domain, compared to hybrid solutions that also employ convolutions. However, Transformers can be surprisingly accurate in low-data regimes.
Future research directions may involve combining X-ray-specific neural modules with the evaluated methods to further improve accuracy.
Stats
The paper utilizes the following key statistics:
The SIXray dataset contains 1,059,231 X-ray images with 8,929 annotated illicit items across 6 classes.
The CFray dataset contains 1,368 annotated X-ray images of parcels with firearms and firearm components across 5 classes.