The 8th "China Software Cup" Competition - Bank Card Number Recognition System
Author: Yiwu Cai, Hongwei Zhu and Kunpeng Ning
- Implemented card number area positioning, card number character recognition;
- Built a fast, easy and user-friendly working interface.
- Provided cross-platform multi-version packages;
- Supported space recognition between characters;
- Collected and enhanced data from complex and multiple scenes.
Use YOLO-v3 network as the basic framework for card number area positioning and character recognition, increase the size of the network input layer and improve the resolution of the image.
When predicting number bounding boxes, multiple prediction frames for a single number may occur. We added NMS between different classes based on original ones.
Due to the privacy of bank cards, only 350 bank card numbers were collected and tagged in real time with limited resources, but the training data was expanded to 14,000 copies through random rotation, panning, background replacement, highlight overlay, brightness/color/blur adjustment...
The list of number box spacing is processed by Softmax, and if it is greater than the specified threshold, We determined there should be a space between two bounding boxes.
Here's a small demo video for anyone interested.