从带噪点纯色白色扫描背景提取照片生成透明图的技术求助

Hey there! I’ve fought this exact battle while digitizing old family photos—scanner dust spots and uneven white backgrounds always mess up basic edge detection. Let’s build a more robust pipeline that ignores those tiny specks and grabs only the actual photo.

核心思路 #

Instead of relying solely on edge detection, we’ll combine noise reduction, adaptive thresholding, and contour filtering to isolate the photo. Here’s the plan:

• Blur the image to suppress small dust specks
• Use adaptive thresholding to handle uneven background lighting
• Apply morphological operations to fill tiny gaps in the photo’s outline
• Filter contours to keep only the largest one (your actual photo)
• Generate a transparent RGBA image using the filtered contour as a mask

分步实现代码 #

import cv2
import numpy as np

# 读取扫描图像
img = cv2.imread("scanned_photo.jpg")
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

# 1. 高斯模糊去噪（过滤小灰尘）
blurred = cv2.GaussianBlur(gray, (5, 5), 0)

# 2. 自适应阈值二值化（处理不均匀背景）
thresh = cv2.adaptiveThreshold(
    blurred, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 11, 2
)

# 3. 形态学闭操作（填充照片边缘的小空洞）
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (7, 7))
closed = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, kernel)

# 4. 寻找并筛选轮廓（只保留最大的轮廓）
contours, _ = cv2.findContours(closed.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
# 按轮廓面积排序，取最大的那个（假设照片是图像中最大的物体）
largest_contour = max(contours, key=cv2.contourArea)

# 创建掩码
mask = np.zeros_like(gray)
cv2.drawContours(mask, [largest_contour], -1, 255, thickness=cv2.FILLED)

# 5. 生成透明背景图像
rgba = cv2.cvtColor(img, cv2.COLOR_BGR2BGRA)
rgba[:, :, 3] = mask  # 设置alpha通道为掩码

# 保存结果
cv2.imwrite("transparent_photo.png", rgba)

关键步骤解释 #

• 高斯模糊: The (5,5) kernel smooths out tiny dust specks without blurring the actual photo edges. Adjust the kernel size if your dust is larger/smaller.
• 自适应阈值: Unlike global thresholding, this adjusts the threshold value locally, which handles uneven white backgrounds common in scans.
• 形态学闭操作: This fills small holes in the photo’s outline caused by dust or minor edge gaps, ensuring a solid mask.
• 轮廓筛选: By picking the largest contour, we ignore all tiny dust-related contours and focus only on the photo itself. If you have multiple photos in one scan, you’ll need to adjust this to filter contours by area range instead.

额外优化建议 #

• If your photo has thin borders that get cut off, try dilating the mask slightly before applying it: dilated_mask = cv2.dilate(mask, kernel, iterations=1)
• For photos with very dark edges, you can tweak the adaptive threshold parameters (like the block size 11 or constant 2) to get a better outline.

内容的提问来源于stack exchange，提问作者user10990260