This repository provides all the dataset with the code used in our paper.
Given input image of Renminbi (RMB) bill, first, the program will perform pre-processing: (i) background removal, and (ii) taking 4 Potential RoIs per two-sides (front-back) RMB bill. Then, given 4 Potential RoIs per each two-sides RMB bill, BPN (Block-wise Prediction Networks) as the region proposer will locate the region of interest (ROI) of bank serial number. The ROI will be separated into individual characters, and finally each individual character will be classified using CNN classifier. (See Figure below)
According the flowchart above, we have two groups of training data: (i) for region proposing, and (ii) for character classification.
(i) For region proposing (can be downloaded here)
- Consist of 20 positive Potential RoIs and 30 negative Potential RoIs,
- Bouding box ground truth (for positive Potential RoIs) is provided in txt file with a format of .
(ii) For character classification (can be downloaded here)
- Consist of 34 classes ("I" and "O" are treated in the same class with "1" and "0", respectively),
- Each class has 20 training images.
The full original data, raw RMB bill images scanned from the machine, with their serial number label can be downloaded here, with specifications as follows.
- Number of images: 2,400 two sides images from 1,200 Renminbi bills,
- Ground truth label: provided in CSV files with 3 columns (1: identifier of the corresponding image filename, 2: serial number (0 means having no serial number), 3: length of serial number).
The code provided here is only for testing version (cannot be used for training), in a single '.py' file: RmbBanknotesRecognition.py
. To run the code:
- First, download the trained model & sample input images here. Input images consist of set of face-back scanned RBM bill images (face-back order can be arbitrary as long as two face-back images from a same RMB bill are placed consecutively),
- Extract and put them in the same folder with the '.py' file,
- Run
RmbBanknotesRecognition.py
(used libraries: Tensoflow, OpenCV, Numpy).
A successful execution will give a result as shown below. Note that the serial numbers in Figures (c) and (d) are intentionally removed and added one character, respectively, from Figure (b).
@inproceedings{umam2018BPN,
title={A Light Deep Learning Based Method for Bank Serial Number Recognition},
author={Umam, Ardian and Chuang, Jen-Hui and Li, Dong-Lin},
booktitle={IEEE International Conference on Visual Communications and Image Processing (VCIP)},
year={2018}
}
The paper can be accessed here.