CNN_Handwritten_Digit_Recognizer (CNN手写数字识别)

A CNN handwritten digit recognizer with graphical UI, implemented with Python (tensorflow, kivy).

一个带界面的CNN手写数字识别，使用Python（tensorflow, kivy）实现。

Introduction (简介)

This is a simple handwritten digit recognizer application, implemented with only Python language. Basically, I use Tensorflow to build the CNN classifier as the recognizer, and use Kivy to realize the UI. This program supports mouse or touch screen, and can recognize 0-9 altogether 10 numbers. The keypoint is as follows:

• Use Tensorflow (as well as its Estimator API) to build the CNN model. Use MNIST dataset to train the model.
• Use Kivy to build the UI.
• After drawing digit on the Kivy App window, the app export the digit as a image, which, after being processed with PIL, is feeded to the CNN model. The recognition result is displayed on the window later.

这是一个简单的手写数字识别小应用，全部使用Python语言实现，主要使用Tensorflow搭建CNN分类器作为后台识别器，使用Kivy框架编写前端界面，支持鼠标或触控操作，能够识别0-9这10个数字。实现的技术要点如下：

• 使用Tensorflow（和其Estimator API）构建CNN模型，使用MNIST数据集训练模型；
• 使用Kivy框架构建前端界面；
• 在Kivy界面上写字后，将界面导出成图片，使用PIL处理后喂给后端的CNN进行识别，识别结果显示在界面上。

Dependencies (依赖环境)

• Python 3
• Tensorflow
• Kivy
• Numpy
• PIL

Run (运行方法)

Run the main app:

python main.py

Train the model (trained model is included, so the main app can work immediately)

python trainer.py

You can also modify some hyperparams in trainer.py.

使用以下命令启动主程序：

python main.py

使用一下命令进行模型的训练（已有训练好的模型，可以直接使用）：

python trainer.py

可以在trainer.py文件中更改各种模型的超参数。

Existing Problems (存在的问题)

Some existing problems, which I will try to find out the reasons and fix in the future when I'm free. There we have:

• When drawing digit at the first time after the app runs, the response time is quite long. I guess it's about the loading of tf, and multi-thread may help.
• The effectiveness of the recognition is not very good. The classifer is easy to mix some digits (like 2 & 7, 2 & 3, 0 & 6 etc.).
• If I modify the x in MNIST dataset, changing values not equal to 0 into 1, the training speed will be much slower. I don't know why.

If you know how to handle these existing problem, or any other problems, feel free to let me know. Thank you!

现在已知存在的问题，之后有时间再找原因修复。问题有：

• 启动程序后，第一次写字，响应速度慢，猜想是第一次要加载tf相关的东西，拖慢了速度，估计可以用多线程解决；
• 识别效果一般，有几个数字（如2和7，2和3，0和6）经常混淆；
• 若将MNIST数据集进行处理，将x中所有非0值改为1，训练速度会极大的下降，不知道是为什么。

存在的这些问题或其他问题，若有大神知道原因或解决方法，求告知。

Finally (最后)

I wrote it just for fun, btw learn the use of Tensorflow and git. I'm new to TF and DL. Any seggestions is welcomed. Thank you!

这个小程序就是自己写着玩，顺便学学Tensorflow和git的使用。TF和深度学习新手，请多包涵，请多指导。

Reference (参考资料)

• 使用tensorflow实现CNN
• Tensorflow-Example
• Tensorflow-Estimator-自定义估算器