MNIST digit recog. using opencv and keras

digit_recog.py

import tensorflow as tf
import cv2
import numpy as np
import matplotlib.pyplot as plt
from tensorflow.keras.models import load_model

dataset = tf.keras.datasets.mnist

#### train - test - split ####
(X_train, y_train), (X_test, y_test) = dataset.load_data()


#### normalize value to b/w 0and1 ###
X_train= X_train/255.0
X_test= X_test/255.0


### CNN (BATCH , HEIGHT, WIDTH, 1)
#### ANN (BATCH_SIZE, FEATURES)
#### FEATURES = WIDTH * HEIGHT
#### reshape array to fit in network ####

X_train = X_train.reshape(X_train.shape[0], -1)
X_test = X_test.reshape(X_test.shape[0], -1)

# (batch_size, height, width, 1)
#### ANN ########

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout

# 0-1
model = Sequential()
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.2))

model.add(Dense(128, activation='relu'))
model.add(Dropout(0.2))

## [0-9] ##
model.add(Dense(10, activation='softmax'))

model.compile('adam', 'sparse_categorical_crossentropy', metrics=['acc'])

model.fit(X_train, y_train, epochs=3, batch_size=12, validation_split=0.1)


#### making prediction #######
plt.imshow(X_test[1255].reshape(28,28), cmap='gray')
plt.xlabel(y_test[1255])
plt.ylabel(np.argmax(model.predict(X_test)[1255]))


model.save('digit_trained.h5')


##### open cv for capture and predicting through camera #####
'''
##### cv2


cap = cv2.VideoCapture(0)
while True:
    ret, img = cap.read()
    #img = cv2.flip(img, 1)
    img = img[200:400, 200:400]
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    _, gray = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY_INV)
    cv2.imshow("gray_wind", gray)
    gray = cv2.resize(gray, (28, 28))
    #cv2.imshow('resized')
    gray = gray.reshape(1, 784)
    result = np.argmax(model.predict(gray))
    result = 'cnn : {}'.format(result)
    cv2.putText(img, org=(25,25), fontFace=cv2.FONT_HERSHEY_SIMPLEX, fontScale=1, text= result, color=(255,0,0), thickness=1)
    cv2.imshow("image", img)
   
    if cv2.waitKey(1) == 13:
        break

cap.release()
cv2.destroyAllWindows()
#plt.imshow(img)
'''


############  prediction via paints ##########
### glob
run = False
ix,iy = -1,-1
follow = 25
img = np.zeros((512,512,1))

### func
def draw(event, x, y, flag, params):
    global run,ix,iy,img,follow
    if event == cv2.EVENT_LBUTTONDOWN:
        run = True
        ix, iy = x, y
    elif event == cv2.EVENT_MOUSEMOVE:
        if run == True:
            cv2.circle(img, (x,y), 20, (255,255,255), -1)

    elif event == cv2.EVENT_LBUTTONUP:
        run = False
        cv2.circle(img, (x,y), 20, (255,255,255), -1)
        gray = cv2.resize(img, (28, 28))
        gray = gray.reshape(1, 784)
        result = np.argmax(model.predict(gray))
        result = 'cnn : {}'.format(result)
        cv2.putText(img, org=(25,follow), fontFace=cv2.FONT_HERSHEY_SIMPLEX, fontScale=1, text= result, color=(255,0,0), thickness=1)
        follow += 25
    elif event == cv2.EVENT_RBUTTONDOWN:
        img = np.zeros((512,512,1))
        follow = 25


### param
cv2.namedWindow('image')
cv2.setMouseCallback('image', draw)



while True:    
    cv2.imshow("image", img)
   
    if cv2.waitKey(1) == 27:
        break

cv2.destroyAllWindows()

########## THANKS ##########

Many times, the paint digit shows error.