OpenCV自动消除三张含人物与风景图像的共同背景技术咨询

Hey there! As someone just starting out with OpenCV, this is a really common (and clever) problem to tackle. Let’s break down your question step by step.

Short answer: Yes, but it has some limitations.

Your core idea—identifying consistent pixels across images to build a background—makes sense, but the wording "保留至少在一张图像中相同的像素点" (keep pixels that are the same in at least one image) needs a tweak. What you actually want is to keep pixels that are consistent across most of the images, since the background is the static element present in all three, while the person is the variable part.

The catch here is edge cases: if the person’s clothing happens to match the background color in a spot on one image, that pixel might get incorrectly labeled as background. Also, if the person doesn’t move much between shots, overlapping pixel regions could be hard to distinguish from the background.

Since you have three images of the same scene, the easiest and most reliable automatic method leverages pixel statistics. The background is stable across all three shots, while the person changes position/appearance. Taking the median value for each pixel position across the three images works perfectly here—median filters out "outlier" pixels (the person) and keeps the consistent background.

Once you have the background, you can subtract it from each original image to isolate the person, then clean up the result with basic post-processing.

Code example #

import cv2
import numpy as np

# Load your three images (replace with your file paths)
img1 = cv2.imread("image1.jpg")
img2 = cv2.imread("image2.jpg")
img3 = cv2.imread("image3.jpg")

# Make sure all images are the same size (resize if needed)
assert img1.shape == img2.shape == img3.shape, "All images must have identical dimensions!"

# Stack the images into a single array (shape: 3, height, width, 3)
stacked_imgs = np.stack([img1, img2, img3], axis=0)

# Calculate median across the three images to get the background
background = np.median(stacked_imgs, axis=0).astype(np.uint8)

# Save the generated background (your Result.jpg)
cv2.imwrite("Result.jpg", background)

# Function to extract foreground (person) from a single image
def get_foreground(original_img, bg_img):
    # Compute difference between original and background
    diff = cv2.absdiff(original_img, bg_img)
    # Convert to grayscale for thresholding
    gray_diff = cv2.cvtColor(diff, cv2.COLOR_BGR2GRAY)
    # Threshold to create a mask of foreground pixels
    _, foreground_mask = cv2.threshold(gray_diff, 25, 255, cv2.THRESH_BINARY)
    # Clean up mask with morphological operations (remove small noise)
    kernel = np.ones((3, 3), np.uint8)
    foreground_mask = cv2.morphologyEx(foreground_mask, cv2.MORPH_OPEN, kernel)
    foreground_mask = cv2.morphologyEx(foreground_mask, cv2.MORPH_CLOSE, kernel)
    # Apply mask to original image to get the person
    foreground = cv2.bitwise_and(original_img, original_img, mask=foreground_mask)
    return foreground

# Process each image to extract the person
fg1 = get_foreground(img1, background)
fg2 = get_foreground(img2, background)
fg3 = get_foreground(img3, background)

# Save your foreground results
cv2.imwrite("person1.jpg", fg1)
cv2.imwrite("person2.jpg", fg2)
cv2.imwrite("person3.jpg", fg3)

Why this works better than your initial idea #

• Fully automatic: No manual marking or extra inputs needed
• Robust to outliers: Median ignores the person’s pixels (which are different across images) far better than averaging, which would blur the background with person pixels
• Easy post-processing: The morphological operations clean up small noise from the difference mask, making the person’s edges look cleaner

If the person moves significantly between shots, you could also try OpenCV’s built-in background subtractors (designed for video, but usable with a few frames):

# Initialize MOG2 background subtractor
bg_subtractor = cv2.createBackgroundSubtractorMOG2(history=3, detectShadows=False)
# Feed each image to the subtractor to build the background model
bg_subtractor.apply(img1)
bg_subtractor.apply(img2)
# Get the foreground mask from the third image
foreground_mask = bg_subtractor.apply(img3)
# Then use the mask to extract the person, same as the get_foreground function above

That said, this method is less reliable with only three frames—background subtractors work best with more data to learn from. The median method is your best bet here.

内容的提问来源于stack exchange，提问作者Borja