Lens Blur effect using Python Pillow

Question

In Pillow, we can blur image using filters:

blurred_image = original_image.filter(ImageFilter.BLUR)

But this is not a true lens effect. Is it possible using some custom filter? Any example?

(Number 2 is what I need. Number 3 is what pillow BLUR filter.)

Answer 1

To simulate the lens blur, you'll have to pay special attention to the bokeh effect. To create the bokeh effect, you'll have to use a binary kernel blurring method (like disc blur where you average all the pixel in around the current pixel in a dish shape). Also you'll need to gamma correct the image so that the dark parts become really dark so that only the bright speckles creates the bokeh. You can then blur the image again without gamma correction and get the max of both.

Pseudo-code:

gammaCorrectedImage = image ^ 3 // Or whatever power works for you.
bokeh = discBlur(gammaCorrectedImage)
bokeh = cubeRoot(bokeh) // To get back the original gamma.
blurImage = discBlur(image)
finalImage = max(bokeh, blurImage)

I got a close enough result:

Answer 2

you need to add depth to the filtering for this task. Lens blur blur more for distances more distant to focused one. I do not expect you got the depth of each pixel so instead:

1. Create masks for each depth layer

   Segmentate the image and then manually chose the depth for each object. For example you should recognize at least these layers:

   • focused object
   • far background
   • near background/foreground

   Then you can also interpolate between the unfocused layers creating smooth gradient between them. Only the focused layer should stay sharp. From all of this create depth map (blur strength) where each pixel will have intensity:

   i=|depth-focused_depth|
   

   Normalized to some usable interval for blurring like float <0.0,1.0> or int <0,256>.

2. use masked blur wiht this depth mask as mask

   see How to blur some portion of Image

   So when put all together I got this:

   

   Where the blur mask was manually (hand) painted then blurred the near and far out of focus parts (gray,black) together leaving focused part (white) without change. Note that in this mask white means no blurring so you need to invert the i

   i(x,y) = white - |depth(x,y)-focused_depth|
   

[Notes]

Step #1 can be exchanged for some predefined masks similarly shaped to object on the image that is been processed as many photos have the focused object centered so the depth masks will be wery similar to each other for similarly shaped objects (like persons).

Answer 3

PIL version of Jack Le's pseudocode for blur with bokeh simulation:

from PIL import ImageFilter

image = Image.open('images/input.png').convert('RGB')
mask = Image.open('images/mask.png').convert('L')
p = 0.4 # power for gamma correction
# r = 3 # in case you want to use Gaussian blur instead of box blur
gammaCorrectedImage = image.point(lambda x: x ** p)  
bokeh = gammaCorrectedImage.filter(ImageFilter.BLUR) #ImageFilter.GaussianBlur(radius=r))  
bokeh = bokeh.point(lambda x: x ** (1/p)) # To get back the original gamma.
bokeh = Image.composite(image, bokeh, mask)
blurImage = image.filter(ImageFilter.BLUR) #ImageFilter.GaussianBlur(radius=r)) 
blurImage = Image.composite(image, blurImage, mask)
finalImage = lighter(bokeh, blurImage) #.getdata(), blurImage.getdata())

finalImage.show()

With Box blur kernel and the following input and mask image below

input image

mask image

obtained the following output with the above code (try changing the parameter p and the blur kernel along with the radius r to get an output closest to the desired effect):

For example, p=0.5 with Gaussian blur kernel with radius r=7 produces the following output image: