crop-slides-from-youtube.py

#!/usr/bin/env python
import numpy
from PIL import ImageGrab, Image


def find_coeffs(pa, pb):
    """Computes the appropiate parameters for PIL.Image.transform
    from the coordinates of two rectangles given as parameters:

    pa: list of four corners of rectangle after transform
    pb: list of four corners of rectangle before transform
    """
    # Adapted from (https://stackoverflow.com/questions/14177744/how-does-perspective-transformation-work-in-pil)
    A = []
    for a, b in zip(pa, pb):  # after, before
        A.append([a[0], a[1], 1, 0, 0, 0, -b[0] * a[0], -b[0] * a[1]])
        A.append([0, 0, 0, a[0], a[1], 1, -b[1] * a[0], -b[1] * a[1]])

    A = numpy.array(A)
    B = numpy.array(pb).reshape(8)

    res = numpy.linalg.solve(A, B)
    return res.reshape(8)


def crop_and_undistort(img, screen_corners=None):
    """Extracts and returns the "slide" part of the image

    Parameters:

    img: input image (frame of video)
    screen_corners: coordinates of NW, NE, SE, SW corners of the
      screen (slide), for an input image which is 1024px wide
    """

    # Get the actual size of img, to adjust scale
    w, _ = img.size

    if not screen_corners:
        # When no screen corners are given, the image is simply
        # rescaled to a width of 1024, so that the user can display
        # this image and find the coordinates of the corners
        img.thumbnail((1024, 1024))
        return img

    # Otherwise, adjust the screen corners received to the scale of the image
    scale = w / 1024
    screen_corners = [(int(x * scale), int(y * scale)) for x, y in screen_corners]

    # Find the approximate aspect ratio of the slide, and use it to
    # compute the height of the output image (the widht will be the
    # same than the input image)
    approx_aspect_ratio = (screen_corners[1][0] - screen_corners[0][0]) / (
        screen_corners[3][1] - screen_corners[0][1]
    )
    new_h = int(w / approx_aspect_ratio)

    # Find the required coefficients for the perspective transform
    coeffs = find_coeffs([(0, 0), (w, 0), (w, new_h), (0, new_h)], screen_corners)

    # Perform the perspective transform
    result = img.transform((w, new_h), Image.PERSPECTIVE, coeffs, Image.BICUBIC)

    return result


# For convenience, some screen corners are already found by
# the author, for some example videos
conferences = {
    "none": None,
    # https://www.youtube.com/watch?v=ySR_FVNX4bQ
    "kubecon": [(0, 0), (702, 48), (702, 418), (0, 447)],

    # https://www.youtube.com/watch?v=QpaapVaL8Fw
    "pydata2015w": [(283, 88), (1013, 73), (1013, 500), (283, 485)],
}


if __name__ == "__main__":
    import sys

    if len(sys.argv) > 1:
        conference = sys.argv[1]
    else:
        conference = "none"

    img = ImageGrab.grabclipboard()
    if img is None:
        print("You should have an image in the clipboard")
        quit()
    if conference.lower() not in conferences:
        print(
            "Sorry, '{}' conference is unknown. Omit the conferece parameter".format(
                conference
            )
        )
        print("and use the output image as base to measure the coordinates")
        print("of the screen, and add them to `conferences` variable in the script")
        quit()
    screen_coordinates = conferences[conference.lower()]
    result = crop_and_undistort(img, screen_coordinates)
    result.save("slide.png")
    result.show()