File size: 4,394 Bytes
5f1ffb3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
from keras.models import load_model
from helpers import resize_to_fit
from imutils import paths
import numpy as np
import imutils
import cv2
import pickle


MODEL_FILENAME = "captcha_model.hdf5"
MODEL_LABELS_FILENAME = "model_labels.dat"
CAPTCHA_IMAGE_FOLDER = "generated_captcha_images"


# Load up the model labels (so we can translate model predictions to actual letters)
with open(MODEL_LABELS_FILENAME, "rb") as f:
    lb = pickle.load(f)

# Load the trained neural network
model = load_model(MODEL_FILENAME)

# Grab some random CAPTCHA images to test against.
# In the real world, you'd replace this section with code to grab a real
# CAPTCHA image from a live website.
captcha_image_files = list(paths.list_images(CAPTCHA_IMAGE_FOLDER))
captcha_image_files = np.random.choice(captcha_image_files, size=(10,), replace=False)

# loop over the image paths
for image_file in captcha_image_files:
    # Load the image and convert it to grayscale
    image = cv2.imread(image_file)
    image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

    # Add some extra padding around the image
    image = cv2.copyMakeBorder(image, 20, 20, 20, 20, cv2.BORDER_REPLICATE)

    # threshold the image (convert it to pure black and white)
    thresh = cv2.threshold(image, 0, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)[1]

    # find the contours (continuous blobs of pixels) the image
    contours = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    # Hack for compatibility with different OpenCV versions
    contours = contours[1] if imutils.is_cv3() else contours[0]

    letter_image_regions = []

    # Now we can loop through each of the four contours and extract the letter
    # inside of each one
    for contour in contours:
        # Get the rectangle that contains the contour
        (x, y, w, h) = cv2.boundingRect(contour)

        # Compare the width and height of the contour to detect letters that
        # are conjoined into one chunk
        if w / h > 1.25:
            # This contour is too wide to be a single letter!
            # Split it in half into two letter regions!
            half_width = int(w / 2)
            letter_image_regions.append((x, y, half_width, h))
            letter_image_regions.append((x + half_width, y, half_width, h))
        else:
            # This is a normal letter by itself
            letter_image_regions.append((x, y, w, h))

    # If we found more or less than 4 letters in the captcha, our letter extraction
    # didn't work correcly. Skip the image instead of saving bad training data!
    if len(letter_image_regions) != 4:
        continue

    # Sort the detected letter images based on the x coordinate to make sure
    # we are processing them from left-to-right so we match the right image
    # with the right letter
    letter_image_regions = sorted(letter_image_regions, key=lambda x: x[0])

    # Create an output image and a list to hold our predicted letters
    output = cv2.merge([image] * 3)
    predictions = []

    # loop over the lektters
    for letter_bounding_box in letter_image_regions:
        # Grab the coordinates of the letter in the image
        x, y, w, h = letter_bounding_box

        # Extract the letter from the original image with a 2-pixel margin around the edge
        letter_image = image[y - 2:y + h + 2, x - 2:x + w + 2]

        # Re-size the letter image to 20x20 pixels to match training data
        letter_image = resize_to_fit(letter_image, 20, 20)

        # Turn the single image into a 4d list of images to make Keras happy
        letter_image = np.expand_dims(letter_image, axis=2)
        letter_image = np.expand_dims(letter_image, axis=0)

        # Ask the neural network to make a prediction
        prediction = model.predict(letter_image)

        # Convert the one-hot-encoded prediction back to a normal letter
        letter = lb.inverse_transform(prediction)[0]
        predictions.append(letter)

        # draw the prediction on the output image
        cv2.rectangle(output, (x - 2, y - 2), (x + w + 4, y + h + 4), (0, 255, 0), 1)
        cv2.putText(output, letter, (x - 5, y - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.55, (0, 255, 0), 2)

    # Print the captcha's text
    captcha_text = "".join(predictions)
    print("CAPTCHA text is: {}".format(captcha_text))

    # Show the annotated image
    cv2.imshow("Output", output)
    cv2.waitKey()