|
|
|
|
|
# High Accuracy Chinese Plate Recognition Framework
|
|
|
|
|
|
|
|
|
|
|
|
### 介绍
|
|
|
|
|
|
This research aims at simply developping plate recognition project on deep learning methods, with low complexity and high speed. This
|
|
|
|
|
|
project has been by some commercial corporations. Free and open source, deploying by Zeusee.
|
|
|
|
|
|
|
|
|
|
|
|
HyperLPR是一个基于Python的使用深度学习针对对中文车牌识别的实现，与开源的[EasyPR](https://github.com/liuruoze/EasyPR)相比，它的检测速度和鲁棒性和多场景的适应性都要好于EasyPR。
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
### 特性
|
|
|
|
|
|
|
|
|
|
|
|
+ 单张720p 识别时间在单核Intel 2.2G CPU(MBP2015 15inch)不低于 140ms。比EasyPR单核识别速度快近10倍左右的时间。
|
|
|
|
|
|
+ 识别率在EasyPR数据集上0-error达到70.2% 1-error识别率达到 89.6%
|
|
|
|
|
|
+ 单线程平均检测时间在EasyPR数据集在保持在160ms以下。基于adaboost检测方法在实时性、召回率、准确率上都不逊于MSER方法。检测recall和easyPR持平。
|
|
|
|
|
|
+ 代码框架轻量，总代码不到1k行。
|
|
|
|
|
|
|
|
|
|
|
|
### 依赖
|
|
|
|
|
|
|
|
|
|
|
|
+ Keras + Theano backend (Tensorflow data order)
|
|
|
|
|
|
+ Theano
|
|
|
|
|
|
+ Numpy
|
|
|
|
|
|
+ Scipy
|
|
|
|
|
|
+ OpenCV
|
|
|
|
|
|
+ scikit-image
|
|
|
|
|
|
|
|
|
|
|
|
### Pipeline
|
|
|
|
|
|
|
|
|
|
|
|
> step1. 使用opencv 的 HAAR cascade 检测车牌大致位置
|
|
|
|
|
|
|
|
|
|
|
|
> step2. Extend 检测到的大致位置的矩形区域
|
|
|
|
|
|
|
|
|
|
|
|
> step3. 使用类似于mser的方式的多级二值化+ransac拟合车牌的上下边界
|
|
|
|
|
|
|
|
|
|
|
|
> step4. 使用CNN regression回归车牌左右边界
|
|
|
|
|
|
|
|
|
|
|
|
> step5. 使用基于纹理场的算法进行车牌校正倾斜
|
|
|
|
|
|
|
|
|
|
|
|
> step6. 使用CNN滑动窗切割字符
|
|
|
|
|
|
|
|
|
|
|
|
> step7. 使用CNN识别字符
|
|
|
|
|
|
|
|
|
|
|
|
### 简单使用方式
|
|
|
|
|
|
|
|
|
|
|
|
```python
|
|
|
|
|
|
from hyperlpr import pipline as pp
|
|
|
|
|
|
import cv2
|
|
|
|
|
|
image = cv2.imread("filename")
|
|
|
|
|
|
image,res = pp.SimpleRecognizePlate(image)
|
|
|
|
|
|
```
|
|
|
|
|
|
### 测试样例
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
### 获取帮助
|
|
|
|
|
|
HyperLPR讨论QQ群：673071218 加前请备注HyperLPR交流。
|
