Deskew Scanned Document

Deskewing document is one of the essential preprcessing step before OCR and HTR. This repo contain proposed method for document deskewing mainly via Fourier Transform.

Techniques

---

• Fourier Transform

Fast Fourier Transform

---

• Non Local Means Denoising

Article about NLMD

Implementation from scratch

---

Requirements

• OpenCV
• Scikit-image
• Numpy

---

Algorithm steps

• Change image to gray format
• Morphological operations
• Fast Fourier Transform
• FNlM Denoising and thresholding of FFT output
• Determines the peaks
• Calculate angle using slope of drawing line

P.S.

This method works well if angle of skew is between -60 and 60 degrees*

---

Vizualisation of each step

---

Code usage

from deskew_fft import RotateDoc
from skimage.transform import rotate
import skimage.io as io
import matplotlib.pyplot as plt

path = 'test_image_1.jpg'
rotate_doc = RotateDoc(
            img_path=path,
            visualize=1,
            resize_ratio=2,
            peak_top_bottom=True,
            synth=False
            )

image = io.imread(path)
ag = rotate_doc.deskewImage()
rotated_image = rotate(image,ag)
if rotate_doc.visualize:
    plt.imshow(rotated_image)
    plt.show()

---

Synthetic test

In this test document manually rotated from -60 to 60 degree and check performance of methdod.

Rotate -60 degree	Deskew 63 degree

Rotate -50 degree	Deskew 53 degree

Rotate -40 degree	Deskew 46 degree

Rotate -30 degree	Deskew 35 degree

Rotate -20 degree	Deskew 25 degree

Rotate -10 degree	Deskew 12 degree

Rotate 10 degree	Deskew -12 degree

Rotate 20 degree	Deskew -24 degree

Rotate 30 degree	Deskew -35 degree

Rotate 40 degree	Deskew -47 degree

Rotate 50 degree	Deskew -55 degree

Rotate 60 degree	Deskew -65 degree

---

---

Synthetic test code

from deskew_fft import RotateDoc
from skimage.transform import rotate
import skimage.io as io
import matplotlib.pyplot as plt

path = 'test_image_1.jpg'
for angle in range(-60,70,10):
    rotated_image = rotate(io.imread(path),angle)
    rotate_doc = RotateDoc(
                            img_path=path,
                            visualize=1,
                            resize_ratio=2,
                            peak_top_bottom=True,
                            synth=True
                            )
    ag = rotate_doc.deskewImage(angle=angle)
    rotated_image = rotate(rotated_image,ag)
    if rotate_doc.visualize:
        plt.imshow(rotated_image)
        plt.show()
    print('-'*100)