 |
 |
 |
 |
 |
 |
| 동준 | 경윤 | 영석 | 태영 | 태성 | 세린 |
| Member | Roles |
|---|---|
| 동준 | Data 직접 검토, K-fold, Dataset Add, Remove dash, Ensemble |
| 영석 | EDA, Data 직접 검토, Streamlit으로 결과 시각화, cv2에서 사용 가능한 이미지 처리 기법 탐색 |
| 경윤 | EDA, Data 직접 검토, augmentation 기법 적용(salt and pepper, binarization, normalize) |
| 태영 | EDA, Data 직접 검토, super resolution, Ensemble |
| 세린 | Data 직접 검토, augmentation 기법 조사 |
| 태성 | 템플릿 코드 작성, Data 직접 검토, Dataset Add |
When performing deep learning tasks, there are typically two main approaches: one focuses on the model, and the other on the data. In this project, we adopted a data-centric approach to tackle an OCR task related to receipts.
This project utilizes a dataset specifically designed for OCR tasks involving receipts. The dataset contains labeled images of various receipt elements, divided into training and test sets
- Training Images: 400
- Training bboxes: 34623
- Test Images: 120
- Lanugages: 4
- Chinese: 100
- Japanese: 100
- thai: 100
- Vietnamese: 100
| Category | Details | Category | Details |
|---|---|---|---|
| Hardware | GPU: V100 32GB × 4 | Python | 3.10 |
| CUDA | 12.1 | PyTorch | 2.1.0 |
| PyTorch Lightning | 1.8.0 | Libraries | Opencv-python(4.10.0.84), numpy(1.24.4) |
| Collaboration Tools | Notion, WandB |
The result of Ensemble (WBF, IoU=0.3):
Super resolution (x4) + Normalize (base) + Remove dash + 3 folds (9:1 train-valid split)
Super resolution (x4) + Normalize (custom) + Remove dash
