- Title
- DeepCornerNet: A Deep Learning Approach for Automated Corner Grading in Trading Cards
- Creator
- Nahar, Lutfun; Islam, Md. Saiful; Awrangjeb, Mohammad; Verhoeve, Rob; Tuxworth, Gervase
- Relation
- 2023 International Conference on Digital Image Computing: Techniques and Applications (DICTA). Proceedings of the 2023 International Conference on Digital Image Computing: Techniques and Applications, DICTA 2023 (Port Macquarie, Australia 28 November - 01 December, 2023) p. 24-31
- Publisher Link
- http://dx.doi.org/10.1109/DICTA60407.2023.00013
- Publisher
- Institute of Electrical and Electronic Engineers (IEEE)
- Resource Type
- conference paper
- Date
- 2023
- Description
- Magic: the Gathering (MTG), Pokemon and other sports cards have a highly lucrative secondary market driven by the value of specific cards and their corresponding quality. Card grading is a crucial assessment process that quantifies the quality and authenticity of trading cards. Graded cards often command higher prices compared to ungraded ones, particularly if they receive a high grade. Professional grading companies like Professional Sports Authenticator and Certified Guaranty Company evaluate cards based on corner, edge, centering, and surface conditions to determine an overall grade. Currently, card grading heavily relies on human visual inspection, which is not only time-consuming but also introduces subjectivity and inconsistencies in the grading process. To address this issue, this paper proposes an automated corner grading system for cards, called DeepCornerNet, using deep learning techniques. We have compared the performance of VGG-16, VGG-19, and InceptionV3 deep neural networks, leveraging transfer learning and freezing layers at different depths to accurately classify various types of corner defects. Additionally, we have addressed the class imbalance problem by employing the focal loss technique. To evaluate the effectiveness of our system, we assembled a real-world dataset consisting of 593 sports cards, provided by our industry partner, which were scanned using an affordable scanner, resulting in a total of 4,744 corner datasets. We have utilised accuracy metrics such as precision, recall, and F1-score. Our experimental results demonstrate that VGG-16 achieves an accuracy of 78% and outperforms other models in terms of precision (77±0.58), recall (74±0.69), and F1-score (74±0.8).
- Subject
- card grading; deep learning; transfer learning; defect prediction; focal loss
- Identifier
- http://hdl.handle.net/1959.13/1502320
- Identifier
- uon:55212
- Identifier
- ISBN:9798350382204
- Language
- eng
- Reviewed
- Hits: 320
- Visitors: 316
- Downloads: 0
Thumbnail | File | Description | Size | Format |
---|