텐서플로 프로젝트) mnist 손글씨 인식을 사용한 숫자인식 그림판

tkinter모듈의 canvas기능을 활용하여

학습 한뒤, 이미지를 불러와 이미 학습된 것에 mnist화 된 이미지를 넣어

결론을 도출할 수 있게 만들었다.

학습을 하는 과정은

드롭아웃을 활용해 1024개짜리 레이어 5개를 거쳐 학습하게 만들었다.

실행사진

코드

import tensorflow as tf
import numpy as np
from PIL import Image, ImageFilter
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("./mnist/data/", one_hot=True)
 
 
def imageprepare(argv):
    im = Image.open(argv).convert('L')
    width = float(im.size[0])
    height = float(im.size[1])
    newImage = Image.new('L', (28, 28), (255))
 
    if width > height:
        nheight = int(round((20.0 / width * height), 0))
        if (nheight == 0):
            nheight = 1
        img = im.resize((20, nheight), Image.ANTIALIAS).filter(ImageFilter.SHARPEN)
        wtop = int(round(((28 - nheight) / 2), 0))
 
        newImage.paste(img, (4, wtop))
    else:
        nwidth = int(round((20.0 / height * width), 0))
        if (nwidth == 0):
            nwidth = 1
        img = im.resize((nwidth, 20), Image.ANTIALIAS).filter(ImageFilter.SHARPEN)
        wleft = int(round(((28 - nwidth) / 2), 0))
        newImage.paste(img, (wleft, 4))
 
 
    tv = list(newImage.getdata())
 
    tva = [(255 - x) * 1.0 / 255.0 for x in tv]
    return tva
 
 
X = tf.placeholder(tf.float32, [None, 784])
Y = tf.placeholder(tf.float32, [None, 10])
keep_prob = tf.placeholder(tf.float32)
 
 
W1 = tf.Variable(tf.random_normal([784, 256], stddev=0.01))
L1 = tf.nn.relu(tf.matmul(X, W1))
L1 = tf.nn.dropout(L1, keep_prob)
 
W2 = tf.Variable(tf.random_normal([256, 256], stddev=0.01))
L2 = tf.nn.relu(tf.matmul(L1, W2))
L2 = tf.nn.dropout(L2, keep_prob)
 
W3 = tf.Variable(tf.random_normal([256, 256], stddev=0.01))
L3 = tf.nn.relu(tf.matmul(X, W1))
L3 = tf.nn.dropout(L1, keep_prob)
 
W4 = tf.Variable(tf.random_normal([256, 256], stddev=0.01))
L4 = tf.nn.relu(tf.matmul(X, W1))
L4 = tf.nn.dropout(L1, keep_prob)
 
W5 = tf.Variable(tf.random_normal([256, 10], stddev=0.01))
model = tf.matmul(L4, W5)
 
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=model, labels=Y))
optimizer = tf.train.AdamOptimizer(0.00001).minimize(cost)
 
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
 
batch_size = 100
total_batch = int(mnist.train.num_examples / batch_size)
 
for epoch in range(10000):
    total_cost = 0
 
    for i in range(total_batch):
        batch_xs, batch_ys = mnist.train.next_batch(batch_size)
 
        _, cost_val = sess.run([optimizer, cost],
                               feed_dict={X: batch_xs,
                                          Y: batch_ys,
                                          keep_prob: 0.8})
        total_cost += cost_val
 
    print('Epoch:', '%04d' % (epoch + 1),
          'Avg. cost =', '{:.3f}'.format(total_cost / total_batch))
 
print('최적화 완료!')
 
is_correct = tf.equal(tf.argmax(model, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(is_correct, tf.float32))
print('정확도:', sess.run(accuracy,
                        feed_dict={X: mnist.test.images,
                                   Y: mnist.test.labels,
                                   keep_prob: 1}))
 
current_num = -1;
def check():
    global current_num
    global textLabel
    x=[imageprepare('./img.png')]
    newArr=[]
    for i in range(784):
        newArr.append(x[0][i])
    newArr2 = []
    newArr2.append(newArr)
 
    labels = sess.run(model,
                      feed_dict={X: newArr2,
                                Y:[[0,0,0,0,0,0,0,0,0,0]],
                                 keep_prob: 1})
    return np.argmax(labels[0])
def checkgo():
    try:
        a = check()
        return a
    except:
        checkgo()
 
 
import tkinter as tk
from PIL import Image,ImageDraw
 
 
class ImageGenerator:
    def __init__(self,parent,posx,posy,*kwargs):
        self.parent = parent
        self.posx = posx
        self.posy = posy
        self.sizex = 200
        self.sizey = 200
        self.b1 = "up"
        self.xold = None
        self.yold = None
        self.drawing_area=tk.Canvas(self.parent,width=self.sizex,height=self.sizey+10)
        self.drawing_area.place(x=self.posx,y=self.posy)
        self.drawing_area.bind("<Motion>", self.motion)
        self.drawing_area.bind("<ButtonPress-1>", self.b1down)
        self.drawing_area.bind("<ButtonRelease-1>", self.b1up)
        self.drawing_area.place(x=self.sizex/7,y=self.sizex/7)
        self.button=tk.Button(self.parent,text="Done!",width=8,command=self.save)
        self.button.place(x=30,y=self.sizey+50)
        self.button1=tk.Button(self.parent,text="Clear!",width=8,command=self.clear)
        self.button1.place(x=135,y=self.sizey+50)
 
        self.text = tk.Text(self.parent,width=10,height=1)
        self.text.insert(tk.INSERT, "예상숫자 : ")
        self.text.pack()
        self.text.place(x=90,y=4)
 
 
 
        self.image=Image.new("RGB",(200,200),(255,255,255))
        self.draw=ImageDraw.Draw(self.image)
 
    def save(self):
        filename = "img.png"
        self.image.save(filename)
        print("예상숫자 : "+ str(checkgo()))
        a = "예상숫자 : "+ str(checkgo())
        self.text.delete('1.0',tk.END)
        self.text.insert(tk.INSERT,a)
 
    def clear(self):
        self.drawing_area.delete("all")
        self.image=Image.new("RGB",(200,200),(255,255,255))
        self.draw=ImageDraw.Draw(self.image)
 
    def b1down(self,event):
        self.b1 = "down"
 
    def b1up(self,event):
        self.b1 = "up"
        self.xold = None
        self.yold = None
 
    def motion(self,event):
        if self.b1 == "down":
            if self.xold is not None and self.yold is not None:
                event.widget.create_line(self.xold,self.yold,event.x,event.y,smooth='true',width=3,fill='blue')
                self.draw.line(((self.xold,self.yold),(event.x,event.y)),(0,128,0),width=10)
 
        self.xold = event.x
        self.yold = event.y
 
if __name__ == "__main__":
    root=tk.Tk()
    root.wm_geometry("%dx%d+%d+%d" % (265, 300, 10, 10))
    root.config(bg='gray')
    ImageGenerator(root,10,10)
    root.mainloop()
 

주인장은 장난이자 실험으로 드롭아웃 5레이어를 만번까지 돌려 보았지만

98.1%의 정확도에서 더이상 올라가지 않았다.(코스트는 0.001까지 봄)

이 코드는 상당히 최적화가 없고, 단순한 기능 구현과 설명만 하기 위해 만든 코드이다.

그대로 쓰진 말고 조금 더 각색해서 쓰길 바란다.