如何使用scikit-image定位图像中加粗编号实现题目图像分割？

Got it, let’s walk through a practical, scikit-image-focused approach to solve this problem—here’s how you can locate those bold, tab-indented number markers (4, 5, 6) and split your image into separate questions:

Step 1: Preprocess the Image to Isolate Text #

First, we need to simplify the image so text (especially bold numbers) stands out clearly. Convert to grayscale and apply binary thresholding to separate text from the background:

from skimage import io, color, filters, measure, morphology
import numpy as np

# Load your target image
img = io.imread("your_question_image.png")
# Convert to grayscale to reduce complexity
gray = color.rgb2gray(img)
# Apply binary thresholding (invert if your text is dark on a light background)
thresh = filters.threshold_otsu(gray)
binary = gray < thresh

# Optional: Clean up small noise with morphological opening
binary_clean = morphology.opening(binary, morphology.square(3))

Step 2: Focus on the Tab-Indented ROI #

Since the numbers are tab-indented, they’ll cluster in a narrow vertical strip on the left side of the image. Crop this region of interest (ROI) to cut down on irrelevant noise:

# Define left ROI (adjust the width percentage to match your image's indentation)
roi_width = int(img.shape[1] * 0.1)  # Use 10% of the image's width as the left ROI
roi = binary_clean[:, :roi_width]

Step 3: Identify Bold Number Regions #

Bold numbers have thicker strokes, so their connected pixel regions will be larger than regular text. Use scikit-image’s region properties to filter these out:

# Label connected pixel regions in the ROI
labeled_roi = measure.label(roi)
# Extract key properties for each region (area, centroid, bounding box)
regions = measure.regionprops(labeled_roi)

# Filter regions to find bold numbers (adjust the area threshold for your font size)
min_bold_area = 100
candidate_regions = []
for region in regions:
    if region.area >= min_bold_area:
        candidate_regions.append({
            "centroid": region.centroid,  # Note: scikit-image uses (y, x) coordinate order
            "bbox": region.bbox  # (min_row, min_col, max_row, max_col)
        })

# Sort candidates by vertical position (since 4 → 5 → 6 appear top to bottom)
candidate_regions.sort(key=lambda x: x["centroid"][0])

Step 4: (Optional) Verify Numbers with OCR #

To make sure we’re only targeting 4, 5, 6 (not other bold elements), add a quick OCR check using Tesseract:

import pytesseract
from skimage import img_as_ubyte

valid_numbers = ["4", "5", "6"]
target_regions = []

for region in candidate_regions:
    min_row, min_col, max_row, max_col = region["bbox"]
    # Extract the region from the grayscale image
    num_img = gray[min_row:max_row, min_col:max_col]
    # Convert to 8-bit format for Tesseract compatibility
    num_img_8bit = img_as_ubyte(num_img)
    # Configure Tesseract to only recognize digits
    text = pytesseract.image_to_string(
        num_img_8bit,
        config="--psm 10 --oem 3 -c tessedit_char_whitelist=0123456789"
    ).strip()
    if text in valid_numbers:
        target_regions.append(region)
        print(f"Found valid number {text} at y-coordinate: {region['centroid'][0]}")

Step 5: Extract Coordinates and Split the Image #

Use the vertical (y) coordinates of your target regions to split the original image into three separate questions:

# Get the y-centroid of each target number (use the top of the bounding box instead if preferred)
split_ys = [int(region["centroid"][0]) for region in target_regions]

# Split the image into individual questions
question_1 = img[:split_ys[1], :]  # From image top to the start of question 5
question_2 = img[split_ys[1]:split_ys[2], :]  # Between question 5 and 6
question_3 = img[split_ys[2]:, :]  # From question 6 to image bottom

# Save the split images
io.imsave("question_1.png", question_1)
io.imsave("question_2.png", question_2)
io.imsave("question_3.png", question_3)

Key Adjustments for Your Image #

• If your image has uneven lighting, replace Otsu thresholding with adaptive thresholding (filters.threshold_local()).
• Tweak min_bold_area and roi_width to match your image’s font size and indentation depth.
• Skip the OCR step if you’re confident in region size/position filtering—you can rely on the vertical ordering of candidates to target 4, 5, 6.

内容的提问来源于stack exchange，提问作者Kaushik