Back to Search
Start Over
Cooperative distillation with X-ray images classifiers for prohibited items detection.
- Source :
-
Engineering Applications of Artificial Intelligence . Jan2024:Part A, Vol. 127, pN.PAG-N.PAG. 1p. - Publication Year :
- 2024
-
Abstract
- As for the characteristics of the objects in the airtight package in the X-ray image, some prohibited items of the airtight package are difficult to be detected from the X-ray images with complex and overlapped backgrounds. In this article, the cooperative knowledge distillation method is used to enhance the prohibited items detection model in the X-ray image. To efficiently implement hard example mining, this article designs an Multi-task Classification Head (MCH) for teachers to provide prior knowledge of image-level and instance-level predictions. Different from the distillation method in which the students imitate the teacher, the algorithm in this article is implemented by the cooperation between teacher and student. In order to verify the effectiveness of this algorithm, a series of related experiments are carried out on PIDray and SIXray respectively. Experiments show that the algorithm improves the AP of State-of-the-Art dense object detectors (e.g., RetinaNet, ATSS, GFL, and TOOD) in SIXray by 1% ∼ 2%. Especially for prohibited items that are difficult to be found in X-ray images, the algorithm is more effective, and the dense object detectors can achieve a performance improvement of about AP of 2% on the Hidden subset of PIDray. The experiments demonstrate that the cooperative knowledge distillation algorithm proposed in this article can effectively enhance the performance of prohibited items detection, particularly showing more pronounced improvements in detecting hard examples. [ABSTRACT FROM AUTHOR]
Details
- Language :
- English
- ISSN :
- 09521976
- Volume :
- 127
- Database :
- Academic Search Index
- Journal :
- Engineering Applications of Artificial Intelligence
- Publication Type :
- Academic Journal
- Accession number :
- 173784953
- Full Text :
- https://doi.org/10.1016/j.engappai.2023.107276