Python OCR/Image Tools/Functions Optimized For Captcha Breaking

📦 Tools/Functions for Image Processing using PIL and/or OpenCV2

from img-prep import *

otsu_thresholding(img)

otsu tresholding

prerequisites:

• gray image

img = cv2.imread('image.png')
gray = toGray(img)
otsu_img = otsu_thresholding(gray)

gaussian_otsu_thresholding(img)

otsu tresholding with gaussian blur

prerequisites:

• gray image

img = cv2.imread('image.png')
gray = toGray(img)
otsu_img = gaussian_otsu_thresholding(gray)

global_thresholding(img)

normal thresholding

prerequisites:

• gray image

img = cv2.imread('image.png')
gray = toGray(img)
final_img = global_thresholding(gray)

bw_image(img)

make image black & white (Binary)

prerequisites:

• gray image

img = cv2.imread('image.png')
gray = toGray(img)
final_img = bw_image(gray)

auto_edge_detection(img, sigma=0.33)

edge detecting with auto lower and upper threshold detection. sigma=0.33 is basicaly best setup for most images.

prerequisites:

• gray image
• slightly blurred

img = cv2.imread('image.png')
gray = toGray(img)
blurred = blur(gray)
final_img = edge_detection(blurred)

wide_edge_detection(img)

edge detecting using wide threshold

prerequisites:

• gray image
• slightly blurred

img = cv2.imread('image.png')
gray = toGray(img)
blurred = blur(gray)
final_img = wide_edge_detection(blurred)

tight_edge_detection(img)

edge detecting using tight threshold

prerequisites:

• gray image
• slightly blurred

img = cv2.imread('image.png')
gray = toGray(img)
blurred = blur(gray)
final_img = tight_edge_detection(blurred)

percentage_scale(img,percentage)

scale image with given percentage keeps aspect ratio

prerequisites:

• none

example, scales 50%

img = cv2.imread('image.png')
scaled = percentage_scale(img, 50)

convolution_2d(img)

Convolves an image with a 5x5 kernel

prerequisites:

• none

img = cv2.imread('image.png')
conv = convolution_2d(img)

laplacian(img)

Laplacian Edge detector

prerequisites:

• gray image
• slightly blurred

img = cv2.imread('image.png')
gray = toGray(img)
blurred = blur(gray)
laplaced = laplacian(blurred)

sobelX(img)

Sobel Edge detector (X) for better edge detection it uses 64Bit in filter and outputs 8Bits

prerequisites:

• none

img = cv2.imread('image.png')
final_img = sobelX(img)

sobelY(img)

Sobel Edge detector (Y) for better edge detection it uses 64Bit in filter and outputs 8Bits

prerequisites:

• none

img = cv2.imread('image.png')
final_img = sobelY(img)

showImg(img)

Shows image using Qt on Windows install Xming to use

prerequisites:

• none

img = cv2.imread('image.png')
showImg(img)

blur(img)

Slightly blur image

prerequisites:

• none

img = cv2.imread('image.png')
blurred = blur(img)

toGray(img)

Convert to grayscale

prerequisites:

• none

img = cv2.imread('image.png')
gray = toGray(img)

FAQ/Explanations

Convolution

Convolution is an operation between every part of an image and an operator (kernel)
or
convolution is simply an element-wise multiplication of two matrices followed by a sum.

Kernel

A kernel is essentially a fixed size array of numerical coefficeints along with an anchor point in that array, which is tipically located at the center.
The functions mostly use a 5x5 kernel using NumPy

kernel = np.ones((5,5),np.float32)/25

or 3x3 kernel defined in python lib

Laplacian (Edge Detection)

Laplacian edge detector uses only one kernel.
It calculates second order derivatives in a single pass
2nd order derivative, so it is extremely sensitive to noise

Original	Result

Sobel (Edge Detection)

Sobel edge detector is a gradient based method based on the first order derivatives.
It calculates the first derivatives of the image separately for the X and Y axes.
Sobel uses 2x 3x3 kernels
first order derivatives

Sobel Kernel X	Sobel Kernel Y

Original

Sobel X	Sobel Y

Tools/Functions for OCR reading