How to Batch Chroma Key Photos (Masking Green Screen)

Question

Goal

I have hundreds of images that all look similar to this one here:

I simply want to use the green screen to create a mask for each image that looks like this one here (the border should preferably be smoothed out a little bit):

Here is the original image if you want to do tests: https://mega.nz/#!0YJnzAJR!GRYI4oNWcsKztHGoK7e4uIv_GvXBjMvyry7cPmyRpRA

---

What I've tried

I found this post where the user used Imagemagick to achieve chroma keying.

for i in *; do convert $i -colorspace HSV -separate +channel \
  \( -clone 0 -background none -fuzz 3% +transparent grey43 \) \
  \( -clone 1 -background none -fuzz 10% -transparent grey100 \) \
  -delete 0,1 -alpha extract -compose Multiply -composite \
  -negate mask_$i; done;

But no matter how I tweak the numbers, the results are not perfect:

---

I feel really dumb, that I cannot find a solution to such a simple problem myself. Also note, that I am using Linux. So no Photoshop or After Effects! :)

But I am sure that there has to be a solution to this problem.

Update 1

I've just tried using this greenscreen script by fmw42 by running ./greenscreen infile.jpg outfile.png and I am rather satisfied with the result. But it takes around 40 seconds to process one image which results in a total 8 hours for all my images (although I have a rather power workstation, see specs below) Maybe this has something to do witch those errors that occur while processing?:

convert-im6.q16: width or height exceeds limit `black' @ error/cache.c/OpenPixelCache/3911.
convert-im6.q16: ImageSequenceRequired `-composite' @ error/mogrify.c/MogrifyImageList/7995.
convert-im6.q16: no images defined `./GREENSCREEN.6799/lut.png' @ error/convert.c/ConvertImageCommand/3258.
convert-im6.q16: unable to open image `./GREENSCREEN.6799/lut.png': No such file or directory @ error/blob.c/OpenBlob/2874.
convert-im6.q16: ImageSequenceRequired `-clut' @ error/mogrify.c/MogrifyImageList/7870.
convert-im6.q16: profile 'icc': 'RGB ': RGB color space not permitted on grayscale PNG `mask.png' @ warning/png.c/MagickPNGWarningHandler/1667.

Workstation specs

• Memory: 125,8 GiB
• Processor: AMD® Ryzen 9 3900x 12-core processor × 24
• Graphics: GeForce GTX 970/PCIe/SSE2 (two of them)

Answer 1

We know that the background is green and is distinguishable from the object by its color, so I suggest using color thresholding. For this, I have written a simple OpenCV Python code to demonstrate the results.

First, we need to install OpenCV.

sudo apt update
pip3 install opencv-python
# verify installation
python3 -c "import cv2; print(cv2.__version__)"

Then, we create a script named skull.py in the same directory with the images.

import cv2
import numpy as np

def show_result(winname, img, wait_time):
    scale = 0.2
    disp_img = cv2.resize(img, None, fx=scale, fy=scale)
    cv2.imshow(winname, disp_img)
    cv2.waitKey(wait_time)

img = cv2.imread('skull.jpg')
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
# define range of green color in HSV
lower_green = np.array([70, 200, 100])
upper_green = np.array([90, 255, 255])
# Threshold the HSV image to extract green color
mask = cv2.inRange(hsv, lower_green, upper_green)
mask = cv2.bitwise_not(mask)

#cv2.imwrite('mask.png', mask)
show_result('mask', mask, 0)
cv2.destroyAllWindows()

You can easily find a tutorial about HSV color operations using OpenCV. I will not go over the functions used here, but one part is important. Image operations are generally done in RGB color space, which holds red, green and blue components. However, HSV is more like human vision system which holds hue, saturation and value components. You can find the conversion here. Since we seperate color based on our perception, HSV is more suitable for this task.

The essential part is to choose the threshold values appropriately. I chose by inspection around 80 for hue (which is max. 180), and above 200 and 100 for saturation and value (max. 255), respectively. You can print the values of a particular pixel by the following lines:

rows,cols,channels = hsv.shape
print(hsv[row, column])

Note that the origin is left upper corner.

Here is the result:

Two things may be needed. One is doing the operation for a set of images, which is trivial using for loops. The other is that if you do not like some portion of the result, you may want to know the pixel location and change the threshold accordingly. This is possible using mouse events.

for i in range(1, 100):
    img = imread(str(i) + '.jpg')

def mouse_callback(event, x, y, flags, params):
    if event == cv2.EVENT_LBUTTONDOWN:
        row = y
        column = x
        print(row, column)

winname = 'img'
cv2.namedWindow(winname)
cv2.setMouseCallback(winname, mouse_callback)

Keep in mind that show_result function resizes the image by scale factor.

If you do not want to deal with pixel positions, rather you want smooth results, you can apply morphological transformations. Especially opening and closing will get the work done.

kernel = np.ones((11,11), np.uint8)
opened = cv2.morphologyEx(mask, cv2.MORPH_OPEN, kernel)
closed = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel)

Result with opening (kernel=11x11):

Answer 2

I can't really fit this in a comment, so I've put it as an answer. If you want to use Fred's greenscreen script, you can hopefully use GNU Parallel to speed it up.

Say you use the commands:

mkdir out
greenscreen image.png out/image.png

to process one image, and you have thousands, you can do the following to keep all your CPU cores busy in parallel till they are all processed:

mkdir out
parallel greenscreen {} out/{} ::: *.png