ルール

正しく認識する
→MNISTの勝ち
正しく認識しない
→私の勝ち（字がヘタって言われている気もしますが。。）

学習モデルの生成

kerasにサンプルコードがあったのでそのまま利用。

mnist_cnn.py
https://github.com/keras-team/keras/blob/master/examples/mnist_cnn.py

生成したモデルを保存する1行だけ追加しました。

model.save('mnist_model.h5')

学習結果はこちら。見方はよくわかりません。99%正解するってことでしょうか。
詳しい人ヘルプ！

  from ._conv import register_converters as _register_converters
Using TensorFlow backend.
x_train shape: (60000, 28, 28, 1)
60000 train samples
10000 test samples
Train on 60000 samples, validate on 10000 samples
Epoch 1/12
60000/60000 [==============================] - 513s 9ms/step - loss: 0.2750 - acc: 0.9163 - val_loss: 0.0586 - val_acc: 0.9806
Epoch 2/12
60000/60000 [==============================] - 559s 9ms/step - loss: 0.0894 - acc: 0.9734 - val_loss: 0.0464 - val_acc: 0.9850
Epoch 3/12
60000/60000 [==============================] - 688s 11ms/step - loss: 0.0658 - acc: 0.9803 - val_loss: 0.0370 - val_acc: 0.9880
Epoch 4/12
60000/60000 [==============================] - 688s 11ms/step - loss: 0.0542 - acc: 0.9834 - val_loss: 0.0309 - val_acc: 0.9895
Epoch 5/12
60000/60000 [==============================] - 693s 12ms/step - loss: 0.0464 - acc: 0.9858 - val_loss: 0.0318 - val_acc: 0.9889
Epoch 6/12
60000/60000 [==============================] - 674s 11ms/step - loss: 0.0417 - acc: 0.9874 - val_loss: 0.0294 - val_acc: 0.9903
Epoch 7/12
60000/60000 [==============================] - 654s 11ms/step - loss: 0.0365 - acc: 0.9889 - val_loss: 0.0292 - val_acc: 0.9900
Epoch 8/12
60000/60000 [==============================] - 665s 11ms/step - loss: 0.0345 - acc: 0.9893 - val_loss: 0.0287 - val_acc: 0.9908
Epoch 9/12
60000/60000 [==============================] - 772s 13ms/step - loss: 0.0322 - acc: 0.9902 - val_loss: 0.0270 - val_acc: 0.9910
Epoch 10/12
60000/60000 [==============================] - 684s 11ms/step - loss: 0.0306 - acc: 0.9907 - val_loss: 0.0295 - val_acc: 0.9908
Epoch 11/12
60000/60000 [==============================] - 748s 12ms/step - loss: 0.0282 - acc: 0.9912 - val_loss: 0.0272 - val_acc: 0.9911
Epoch 12/12
60000/60000 [==============================] - 657s 11ms/step - loss: 0.0259 - acc: 0.9918 - val_loss: 0.0252 - val_acc: 0.9916
Test loss: 0.025242764521988738
Test accuracy: 0.9916

予測

以下のコードにて、予測させました。features.argmax()を予測した文字としてます。

# coding:utf-8

import keras
import numpy as np
from keras.models import load_model
from keras.preprocessing.image import array_to_img, img_to_array,load_img
import os
import re

model = load_model('mnist_model.h5')

def list_pictures(directory, ext='jpg|jpeg|bmp|png|ppm'):
    return [os.path.join(root, f)
            for root, _, files in os.walk(directory) for f in files
            if re.match(r'([\w]+\.(?:' + ext + '))', f.lower())]

for picture in list_pictures('./tegaki/'):
    X = []
    img = img_to_array(
        load_img(picture, target_size=(28, 28), grayscale=True))
    X.append(img)

    X = np.asarray(X)
    X = X.astype('float32')
    X = X / 255.0

    features = model.predict(X)

    print('----------')
    print(picture)
    print(features.argmax())
    print('----------')