How to generate a paper-like background with OpenCV

Question

I'm trying to implement a random paper-like background using OpenCV. The attached example shows the real paper document artifacts (please ignore handwriting on it).

A simple effect of raw paper is easily possible to achieve just by applying noise

import cv2
import numpy as np

BG_COLOR = 209

def blank_image(width=1024, height=1024):
    img = np.full((height, width, 1), BG_COLOR, np.uint8)
    return img

def noisy(image):
    row, col, ch = image.shape
    mean = 0
    sigma = 10
    gauss = np.random.normal(mean, sigma, (row, col, ch))
    gauss = gauss.reshape(row, col, ch)
    noisy = gauss + image
    return noisy

if __name__ == '__main__':
    img = blank_image()
    cv2.imwrite('out.jpg', noisy(img))

But it looks too uniform without artifacts above:

And I wonder what are the best approaches of generating such random structures from the first picture.

Answer 1

Inspired with the solution provided by BoboDarph, I created a paper-like texture. Now I need to add artifacts from the real photo above. The code is below:

import cv2
import numpy as np


BG_COLOR = 209
BG_SIGMA = 5
MONOCHROME = 1


def blank_image(width=1024, height=1024, background=BG_COLOR):
    """
    It creates a blank image of the given background color
    """
    img = np.full((height, width, MONOCHROME), background, np.uint8)
    return img


def add_noise(img, sigma=BG_SIGMA):
    """
    Adds noise to the existing image
    """
    width, height, ch = img.shape
    n = noise(width, height, sigma=sigma)
    img = img + n
    return img.clip(0, 255)


def noise(width, height, ratio=1, sigma=BG_SIGMA):
    """
    The function generates an image, filled with gaussian nose. If ratio parameter is specified,
    noise will be generated for a lesser image and then it will be upscaled to the original size.
    In that case noise will generate larger square patterns. To avoid multiple lines, the upscale
    uses interpolation.

    :param ratio: the size of generated noise "pixels"
    :param sigma: defines bounds of noise fluctuations
    """
    mean = 0
    assert width % ratio == 0, "Can't scale image with of size {} and ratio {}".format(width, ratio)
    assert height % ratio == 0, "Can't scale image with of size {} and ratio {}".format(height, ratio)

    h = int(height / ratio)
    w = int(width / ratio)

    result = np.random.normal(mean, sigma, (w, h, MONOCHROME))
    if ratio > 1:
        result = cv2.resize(result, dsize=(width, height), interpolation=cv2.INTER_LINEAR)
    return result.reshape((width, height, MONOCHROME))


def texture(image, sigma=BG_SIGMA, turbulence=2):
    """
    Consequently applies noise patterns to the original image from big to small.

    sigma: defines bounds of noise fluctuations
    turbulence: defines how quickly big patterns will be replaced with the small ones. The lower
    value - the more iterations will be performed during texture generation.
    """
    result = image.astype(float)
    cols, rows, ch = image.shape
    ratio = cols
    while not ratio == 1:
        result += noise(cols, rows, ratio, sigma=sigma)
        ratio = (ratio // turbulence) or 1
    cut = np.clip(result, 0, 255)
    return cut.astype(np.uint8)


if __name__ == '__main__':
    cv2.imwrite('texture.jpg', texture(blank_image(background=230), sigma=4, turbulence=4))
    cv2.imwrite('texture-and-noise.jpg', add_noise(texture(blank_image(background=230), sigma=4), sigma=10))

    cv2.imwrite('noise.jpg', add_noise(blank_image(1024, 1024), sigma=10))

The resulting picture: