Image shadow and highlight correction/adujstment with opencv.

ShadowHighlightCorrection.py

import numpy as np
import cv2

def correction(
        img,
        shadow_amount_percent, shadow_tone_percent, shadow_radius,
        highlight_amount_percent, highlight_tone_percent, highlight_radius,
        color_percent
):
    """
    Image Shadow / Highlight Correction. The same function as it in Photoshop / GIMP
    :param img: input RGB image numpy array of shape (height, width, 3)
    :param shadow_amount_percent [0.0 ~ 1.0]: Controls (separately for the highlight and shadow values in the image) how much of a correction to make.
    :param shadow_tone_percent [0.0 ~ 1.0]: Controls the range of tones in the shadows or highlights that are modified.
    :param shadow_radius [>0]: Controls the size of the local neighborhood around each pixel
    :param highlight_amount_percent [0.0 ~ 1.0]: Controls (separately for the highlight and shadow values in the image) how much of a correction to make.
    :param highlight_tone_percent [0.0 ~ 1.0]: Controls the range of tones in the shadows or highlights that are modified.
    :param highlight_radius [>0]: Controls the size of the local neighborhood around each pixel
    :param color_percent [-1.0 ~ 1.0]:
    :return:
    """
    shadow_tone = shadow_tone_percent * 255
    highlight_tone = 255 - highlight_tone_percent * 255

    shadow_gain = 1 + shadow_amount_percent * 6
    highlight_gain = 1 + highlight_amount_percent * 6

    # extract RGB channel
    height, width = img.shape[:2]
    img = img.astype(np.float)
    img_R, img_G, img_B = img[..., 2].reshape(-1), img[..., 1].reshape(-1), img[..., 0].reshape(-1)

    # The entire correction process is carried out in YUV space,
    # adjust highlights/shadows in Y space, and adjust colors in UV space
    # convert to Y channel (grey intensity) and UV channel (color)
    img_Y = .3 * img_R + .59 * img_G + .11 * img_B
    img_U = -img_R * .168736 - img_G * .331264 + img_B * .5
    img_V = img_R * .5 - img_G * .418688 - img_B * .081312

    # extract shadow / highlight
    shadow_map = 255 - img_Y * 255 / shadow_tone
    shadow_map[np.where(img_Y >= shadow_tone)] = 0
    highlight_map = 255 - (255 - img_Y) * 255 / (255 - highlight_tone)
    highlight_map[np.where(img_Y <= highlight_tone)] = 0

    # // Gaussian blur on tone map, for smoother transition
    if shadow_amount_percent * shadow_radius > 0:
        # shadow_map = cv2.GaussianBlur(shadow_map.reshape(height, width), ksize=(shadow_radius, shadow_radius), sigmaX=0).reshape(-1)
        shadow_map = cv2.blur(shadow_map.reshape(height, width), ksize=(shadow_radius, shadow_radius)).reshape(-1)

    if highlight_amount_percent * highlight_radius > 0:
        # highlight_map = cv2.GaussianBlur(highlight_map.reshape(height, width), ksize=(highlight_radius, highlight_radius), sigmaX=0).reshape(-1)
        highlight_map = cv2.blur(highlight_map.reshape(height, width), ksize=(highlight_radius, highlight_radius)).reshape(-1)

    # Tone LUT
    t = np.arange(256)
    LUT_shadow = (1 - np.power(1 - t * (1 / 255), shadow_gain)) * 255
    LUT_shadow = np.maximum(0, np.minimum(255, np.int_(LUT_shadow + .5)))
    LUT_highlight = np.power(t * (1 / 255), highlight_gain) * 255
    LUT_highlight = np.maximum(0, np.minimum(255, np.int_(LUT_highlight + .5)))

    # adjust tone
    shadow_map = shadow_map * (1 / 255)
    highlight_map = highlight_map * (1 / 255)

    iH = (1 - shadow_map) * img_Y + shadow_map * LUT_shadow[np.int_(img_Y)]
    iH = (1 - highlight_map) * iH + highlight_map * LUT_highlight[np.int_(iH)]
    img_Y = iH

    # adjust color
    if color_percent != 0:
        # color LUT
        if color_percent > 0:
            LUT = (1 - np.sqrt(np.arange(32768)) * (1 / 128)) * color_percent + 1
        else:
            LUT = np.sqrt(np.arange(32768)) * (1 / 128) * color_percent + 1

        # adjust color saturation adaptively according to highlights/shadows
        color_gain = LUT[np.int_(img_U ** 2 + img_V ** 2 + .5)]
        w = 1 - np.minimum(2 - (shadow_map + highlight_map), 1)
        img_U = w * img_U + (1 - w) * img_U * color_gain
        img_V = w * img_V + (1 - w) * img_V * color_gain

    # re convert to RGB channel
    output_R = np.int_(img_Y + 1.402 * img_V + .5)
    output_G = np.int_(img_Y - .34414 * img_U - .71414 * img_V + .5)
    output_B = np.int_(img_Y + 1.772 * img_U + .5)

    output = np.row_stack([output_B, output_G, output_R]).T.reshape(height, width, 3)
    output = np.minimum(output, 255).astype(np.uint8)
    return output

Thanks for sharing this great code snippet! I found the code may give very saturated spots when adjusting the saturation. After some debugging I found this is because Line 90 didn't do a guarding on the lower (0) side. If I change it to output = np.minimum(np.maximum(output, 0), 255).astype(np.uint8), the spots would disappear. Just FYI. (Not sure how to submit a PR for a gist)

Thanks for sharing this piece of code @HViktorTsoi !
May I know what does the parameter 'color_percent' exactly affect ?

This was very helpful, thank you!

VERY GOOD. THANKS

Thanks for sharing this great code snippet! I found the code may give very saturated spots when adjusting the saturation. After some debugging I found this is because Line 90 didn't do a guarding on the lower (0) side. If I change it to output = np.minimum(np.maximum(output, 0), 255).astype(np.uint8), the spots would disappear. Just FYI. (Not sure how to submit a PR for a gist)

I have the same problem and your solution solved it. Many thanks bro 100

Thanks for sharing this piece of code @HViktorTsoi ! May I know what does the parameter 'color_percent' exactly affect ?

I think w = 1 - np.minimum(2 - (shadow_map + highlight_map), 1) should be changed to "w = np.minimum(2 - (shadow_map + highlight_map), 1)".

Thanks a lot for sharing the code, it does its job perfectly. Although, there is one problem - it darkens highlights and lightens shadows, but it doesn't seems able to do the inverse - increase and darken them respectively. I'd like to know how this function can be modified for the sake of achieving that effect. And if it's already possible - how the inputs should be adjusted.

If you want to add rectangles etc onto the image afterwards you get the following error:
cv2.error: OpenCV: (-5:Bad argument) in function 'rectangle'

Making a copy of the image upon return on line 90 will fix this:

output = np.minimum(output, 255).astype(np.uint8).copy()
or if you changed it as per @gapeot suggestion:
output = np.minimum(np.maximum(output, 0), 255).astype(np.uint8).copy()