| Paper Abstract and Keywords |
| Presentation |
2024-10-03 14:50
Accuracy Improvement of Real Image Classification Using Synthetic Images for Training by Bilateral Filtering
Masakazu Ohkoba, Takeru Inoue, Miho Adachi, Junya Morioka (Meiji Univ.), Kouji Gakuta, Etsuji Yamada, Aoi Kariya (DPS), Masakazu Kinosada, Yujiro Kitaide (Shinsei Printing Co., Ltd.), Ryusuke Miyamoto (Meiji Univ.) |
| Abstract |
(in Japanese) |
(See Japanese page) |
| (in English) |
In the development of machine-learning-based component classification applications, creating a suitable training dataset to achieve practical accuracy requires considerable costs. To reduce this cost,we adopt a dataset generation method using three-dimensional models.The accuracy of classifiers trained on CG datasets decreases when applied to real-world images due to textural differences. To address this domain gap,we propose a method that uses bilateral filter to reduce the visual discrepancy between CG data and real images while preserving edges. By applying bilateral filter to real images, we improved the classification accuracy from 95.08% (without filter) to 96.20% (with filter). Furthermore, by applying bilateral filter to the training dataset,we achieved an accuracy of 96.42%. |
| Keyword |
(in Japanese) |
(See Japanese page) |
| (in English) |
deep learning / CG-based dataset / bilateral filter / domain gap / / / / |
| Reference Info. |
ITE Tech. Rep. |
| Paper # |
|
| Date of Issue |
|
| ISSN |
Online edition: ISSN 2424-1970 |
| Download PDF |
|
|