دور زدن کپچا با شبکه عصبی پیچشی در پایتون — راهنمای گام به گام

1import pandas as pd
2import numpy as np
3import cv2
4import glob
5import imutils
6from imutils import paths
7import os
8import os.path
9
10captcha_image = '11114.png'
11# Load the image and convert it to grayscale
12image = cv2.imread(captcha_image)
13gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
14 
15# grab the base filename as the text
16filename = os.path.basename(captcha_image)
17captcha_text = os.path.splitext(filename)[0]

1# Adding some extra padding around the image
2gray = cv2.copyMakeBorder(gray, 8, 8, 8, 8, cv2.BORDER_REPLICATE)
3
4# applying threshold
5thresh = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY_INV, cv2.THRESH_OTSU)[1]

1# creating empty list for holding the coordinates of the letters
2letter_image_regions = []
3 
4# finding the contours
5contours, hierarchy = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
6for contour in contours:
7  # Get the rectangle that contains the contour
8  (x, y, w, h) = cv2.boundingRect(contour)
9        
10  # checking if any counter is too wide
11  # if countour is too wide then there could be two letters joined together or are very close to each other
12  if w / h > 1.25:
13    # Split it in half into two letter regions
14    half_width = int(w / 2)
15    letter_image_regions.append((x, y, half_width, h))
16    letter_image_regions.append((x + half_width, y, half_width, h))
17  else:  
18    letter_image_regions.append((x, y, w, h))

1letter_image_regions = sorted(letter_image_regions, key=lambda x: x[0])
2
3# Save each letter as a single image
4for letter_bounding_box, letter_text in zip(letter_image_regions, captcha_text):
5  # Grab the coordinates of the letter in the image
6  x, y, w, h = letter_bounding_box
7
8  # Extract the letter from the original image with a 2-pixel margin around the edge
9  letter_image = gray[y - 2:y + h + 2, x - 2:x + w + 2]

1SOLVED_CAPTCHA_FOLDER = "solved-captchas"
2OUTPUT_FOLDER = "extracted_letters"
3
4
5# Get the path of all the solved captcha images
6solved_captchas = glob.glob(os.path.join(SOLVED_CAPTCHA_FOLDER, "*"))
7counts = {}
8
9# loop over the image paths
10for (i, captcha) in enumerate(solved_captchas):
11    print("processing image {}/{}".format(i + 1, len(solved_captchas)))
12
13    
14    # grab the base filename as the text
15    filename = os.path.basename(captcha)
16    captcha_text = os.path.splitext(filename)[0]
17
18    # Load the image and convert it to grayscale
19    image = cv2.imread(captcha)
20    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
21    
22    # Add some extra padding around the image
23    gray = cv2.copyMakeBorder(gray, 8, 8, 8, 8, cv2.BORDER_REPLICATE)
24
25    # applying threshold
26    thresh = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY_INV, cv2.THRESH_OTSU)[1]
27
28     # finding the contours
29    contours, hierarchy = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
30
31    # creating empty list for holding the coordinates of the letters
32    letter_image_regions = []
33
34    # Now we will loop through each of the contours and extract the letter
35    for contour in contours:
36        # Get the rectangle that contains the contour
37        (x, y, w, h) = cv2.boundingRect(contour)
38        
39        # checking if any counter is too wide
40        # if countour is too wide then there could be two letters joined together or are very close to each other
41        if w / h > 1.25:
42            # Split it in half into two letter regions
43            half_width = int(w / 2)
44            letter_image_regions.append((x, y, half_width, h))
45            letter_image_regions.append((x + half_width, y, half_width, h))
46        else:
47            
48            letter_image_regions.append((x, y, w, h))
49            
50
51    
52    # Sort the detected letter images based on the x coordinate to make sure
53    # we get them from left-to-right so that we match the right image with the right letter
54    letter_image_regions = sorted(letter_image_regions, key=lambda x: x[0])
55    
56    # Save each letter as a single image
57    for letter_bounding_box, letter_text in zip(letter_image_regions, captcha_text):
58        # Grab the coordinates of the letter in the image
59        x, y, w, h = letter_bounding_box
60
61        # Extract the letter from the original image with a 2-pixel margin around the edge
62        letter_image = gray[y - 2:y + h + 2, x - 2:x + w + 2]
63
64
65        # Get the folder to save the image in
66        save_path = os.path.join(OUTPUT_FOLDER, letter_text)
67
68        # creating different output folder for storing different letters
69        if not os.path.exists(save_path):
70            os.makedirs(save_path)
71
72        # write the letter image to a file
73        count = counts.get(letter_text, 1)
74        p = os.path.join(save_path, "{}.png".format(str(count)))
75        cv2.imwrite(p, letter_image)
76
77        # increment the count
78        counts[letter_text] = count + 1

1letter_folder = 'extracted_letters'
2
3#creating empty lists for storing image data and labels
4data = []
5labels = []
6for image in paths.list_images(letter_folder):
7    img = cv2.imread(image)
8    img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
9    img = cv2.resize(img, (30,30))
10    
11    # adding a 3rd dimension to the image
12    img = np.expand_dims(img, axis = 2)
13    
14    #grabing the name of the letter based on the folder it is present in
15    label = image.split(os.path.sep)[-2]
16    
17    # appending to the empty lists
18    data.append(img)
19    labels.append(label)
20
21#converting data and labels to np array
22data = np.array(data, dtype = "float")
23labels = np.array(labels)

1#scaling the values of  data between 0 and 1
2data = data/255.0
3
4# Split the training data into separate train and test sets
5from sklearn.model_selection import train_test_split
6(train_x, val_x, train_y, val_y) = train_test_split(data, labels, test_size=0.2, random_state=0)
7
8#one hot encoding
9from sklearn.preprocessing import LabelBinarizer
10lb = LabelBinarizer().fit(train_y)
11train_y = lb.transform(train_y)
12val_y = lb.transform(val_y)

1from keras.models import Sequential
2from keras.layers.convolutional import Conv2D, MaxPooling2D
3from keras.layers.core import Flatten, Dense, Dropout
4from keras.callbacks import EarlyStopping
5
6#building model
7model = Sequential()
8model.add(Conv2D(20, (5, 5), padding="same", input_shape=(30, 30, 1), activation="relu"))
9model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
10model.add(Conv2D(50, (5, 5), padding="same", activation="relu"))
11model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
12model.add(Flatten())
13model.add(Dense(128, activation="relu"))
14model.add(Dropout(0.3))
15model.add(Dense(9, activation="softmax"))
16
17model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])

1# using early stoping for avoiding overfitting
2estop = EarlyStopping(patience=10, mode='min', min_delta=0.001, monitor='val_loss')
3
4model.fit(train_x, train_y, validation_data=(val_x, val_y), batch_size=32, epochs=50, verbose=1, callbacks = [estop])

1# Load the image and convert it to grayscale
2image = cv2.imread('unsolved-captchas.png')
3gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
4    
5# Add some extra padding around the image
6gray = cv2.copyMakeBorder(gray, 8, 8, 8, 8, cv2.BORDER_REPLICATE)
7
8# threshold the image
9thresh = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY_INV, cv2.THRESH_OTSU)[1]
10
11# find the contours
12contours, hierarchy = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
13
14
15    
16letter_image_regions = []
17
18# Now we can loop through each of the contours and extract the letter
19
20for contour in contours:
21    # Get the rectangle that contains the contour
22    (x, y, w, h) = cv2.boundingRect(contour)
23    
24    # checking if any counter is too wide
25    # if countour is too wide then there could be two letters joined together or are very close to each other
26    if w / h > 1.25:
27        # Split it in half into two letter regions
28        half_width = int(w / 2)
29        letter_image_regions.append((x, y, half_width, h))
30        letter_image_regions.append((x + half_width, y, half_width, h))
31    else:
32        letter_image_regions.append((x, y, w, h))
33            
34
35# Sort the detected letter images based on the x coordinate to make sure
36# we get them from left-to-right so that we match the right image with the right letter  
37
38letter_image_regions = sorted(letter_image_regions, key=lambda x: x[0])
39
40# Create an output image and a list to hold our predicted letters
41output = cv2.merge([gray] * 3)
42predictions = []
43    
44# Creating an empty list for storing predicted letters
45predictions = []
46    
47# Save out each letter as a single image
48for letter_bounding_box in letter_image_regions:
49    # Grab the coordinates of the letter in the image
50    x, y, w, h = letter_bounding_box
51
52    # Extract the letter from the original image with a 2-pixel margin around the edge
53    letter_image = gray[y - 2:y + h + 2, x - 2:x + w + 2]
54
55    letter_image = cv2.resize(letter_image, (30,30))
56        
57    # Turn the single image into a 4d list of images
58    letter_image = np.expand_dims(letter_image, axis=2)
59    letter_image = np.expand_dims(letter_image, axis=0)
60
61    # making prediction
62    pred = model.predict(letter_image)
63        
64    # Convert the one-hot-encoded prediction back to a normal letter
65    letter = lb.inverse_transform(pred)[0]
66    predictions.append(letter)
67
68
69    # draw the prediction on the output image
70    cv2.rectangle(output, (x - 2, y - 2), (x + w + 4, y + h + 4), (0, 255, 0), 1)
71    cv2.putText(output, letter, (x - 5, y - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.55, (0, 255, 0), 2)
72
73# Print the captcha's text
74captcha_text = "".join(predictions)
75print("CAPTCHA text is: {}".format(captcha_text))
76
77# Show the annotated image
78cv2.imshow('output',output)
79cv2.waitKey(0)