Table 1.
Summary of the state-of-the-art baggage threat detection frameworks *.
| Literature | Methodology | Performance | Limitations |
|---|---|---|---|
| Miao et al. [14] | Developed CHR [14], an imbalanced resistant framework that leverages reversed connections class-balanced loss function to effectively learn the imbalanced suspicious item categories in a highly imbalanced SIXray [14] dataset. | Achieved an overall mean average precision score of 0.793, 0.606, and 0.381 on SIXray10, SIXray100, and SIXray1000 [14], respectively when coupled with ResNet-101 [46] for recognizing five suspicious item categories. | Although the framework is resistant to an imbalanced dataset, it is still tested only on a single dataset. |
| Hassan et al. [11] | Proposed a contour instance segmentation framework for recognizing baggage threats regardless of the scanner specifications. | Achieved a mean average precision score of 0.4657 on a total of 223,686 multivendor baggage X-ray scans. | Built upon a conventional fine-tuning approach that requires a large-scale training dataset. |
| Gaus et al. [51] | Evaluated the transferability of different one-staged and two-staged object detection and instance segmentation models on SIXray10 [14] subset of the SIXray [14] dataset and also on their locally prepared dataset. | Achieved a mean average precision of 0.8500 for extracting guns and knives on SIXray10 [14] dataset. | Tested on only one public dataset i.e., the SIXray10 [14] for only extracting guns and knives. |
| Wei et al. [13] | Proposed a plug-and-play module dubbed DOAM [13] that can be integrated with the deep object detectors to recognize and localized the occluded threatening items. | Achieved the mean average precision score of 0.740 coupled with SSD [54]. | DOAM [13] is not tested on publicly available GDXray [15] and SIXray [14] datasets. |
| Hassan et al. [5] | Developed a CST framework that leverages contours of the baggage content to generate object proposals that are screened via a single classification backbone. | Achieved a mean average precision score of 0.9343 and 0.9595 on GDXray [15] and SIXray [14] datasets. | CST, although, is tested on two public datasets, but it requires extensive parameter tuning to work well on both of them. |
* For a detailed overview on the existing approaches, we refer the reader to the Supplementary Material of this article.