멈춤보단 천천히라도

import tensorflow as tf


mnist = tf.keras.datasets.mnist

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

model = tf.keras.models.Sequential([
  tf.keras.layers.Flatten(),
  tf.keras.layers.Dense(512, activation=tf.nn.relu),
  tf.keras.layers.Dropout(0.2),
  tf.keras.layers.Dense(10, activation=tf.nn.softmax)
])

model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

model.fit(x_train, y_train, epochs=5)
model.evaluate(x_test, y_test)

model.save_weights('mnist_checkpoint')

import tensorflow as tf
import cv2
import numpy as np
import math


def process(img_input):

    gray = cv2.cvtColor(img_input, cv2.COLOR_BGR2GRAY)

    gray = cv2.resize(gray, (28, 28), interpolation=cv2.INTER_AREA)


    (thresh, img_binary) = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)


    h,w = img_binary.shape

 
    ratio = 100/h
    new_h = 100
    new_w = w * ratio

    img_empty = np.zeros((110,110), dtype=img_binary.dtype)
    img_binary = cv2.resize(img_binary, (int(new_w), int(new_h)), interpolation=cv2.INTER_AREA)
    img_empty[:img_binary.shape[0], :img_binary.shape[1]] = img_binary

    img_binary = img_empty


    cnts = cv2.findContours(img_binary.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)


    # 컨투어의 무게중심 좌표를 구합니다.
    M = cv2.moments(cnts[0][0])
    center_x = (M["m10"] / M["m00"])
    center_y = (M["m01"] / M["m00"])

    # 무게 중심이 이미지 중심으로 오도록 이동시킵니다.
    height,width = img_binary.shape[:2]
    shiftx = width/2-center_x
    shifty = height/2-center_y

    Translation_Matrix = np.float32([[1, 0, shiftx],[0, 1, shifty]])
    img_binary = cv2.warpAffine(img_binary, Translation_Matrix, (width,height))


    img_binary = cv2.resize(img_binary, (28, 28), interpolation=cv2.INTER_AREA)
    flatten = img_binary.flatten() / 255.0

    return flatten



model = tf.keras.models.Sequential([
  tf.keras.layers.Flatten(),
  tf.keras.layers.Dense(512, activation=tf.nn.relu),
  tf.keras.layers.Dropout(0.2),
  tf.keras.layers.Dense(10, activation=tf.nn.softmax)
])

model.load_weights('mnist_checkpoint')


cap = cv2.VideoCapture(0)
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))


while(True):

    ret, img_color = cap.read()

    if ret == False:
        break;

    img_input = img_color.copy()
    cv2.rectangle(img_color, (250, 150),  (width-250, height-150), (0, 0, 255), 3)
    cv2.imshow('bgr', img_color)

    img_roi = img_input[150:height-150, 250:width-250]


    key = cv2.waitKey(1)

    if key == 27:
        break
    elif key == 32:
        flatten = process(img_roi)
       
        predictions = model.predict(flatten[np.newaxis,:])

        with tf.compat.v1.Session() as sess:
            print(tf.argmax(predictions, 1).eval())

        cv2.imshow('img_roi', img_roi)
        cv2.waitKey(0)


cap.release()
cv2.destroyAllWindows()