OpenCV清晰物体边界检测失败：文档边缘检测方案求助

问题背景 #

待检测原图：

尝试四种OpenCV方案（早期方案已注释）后，始终得到错误边缘检测结果：

用户原有代码：

def detectDocEdge(imagePath):
    image = loadImage(imagePath)
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    
    # Approach 1
    # equalized_image = cv2.equalizeHist(gray_image)
    # blurred = cv2.GaussianBlur(equalized_image, (15, 15), 0)  # TODO: May not be required
    # edges = cv2.Canny(blurred, 5, 150, apertureSize=3)
    # kernel = np.ones((15, 15), np.uint8)
    # dilated_edges = cv2.dilate(edges, kernel, iterations=1)

    # Approach 2
    # blurred = cv2.GaussianBlur(gray_image, (5, 5), 0)
    # thresh = cv2.adaptiveThreshold(blurred, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 11, 2)
    # kernel = np.ones((5, 5), np.uint8)
    # dilated_edges = cv2.dilate(thresh, kernel, iterations=1)

    # Approach 3
    # Apply morphological operations
    # kernel = np.ones((3, 3), np.uint8)
    # closed_image = cv2.morphologyEx(thresholded_image, cv2.MORPH_CLOSE, kernel)

    # Approach 4
    blurred = cv2.GaussianBlur(gray_image, (5, 5), 0)
    edges = cv2.Canny(blurred, 50, 150)
    dilated = cv2.dilate(edges, None, iterations=5)
    eroded = cv2.erode(dilated, None, iterations=5)

    contours, hierarchy = cv2.findContours(eroded, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

    contour_img = image.copy()
    cv2.drawContours(contour_img, contours, -1, (0, 255, 0), 2)
    showImage(contour_img)

    max_contour = None
    max_measure = 0  # This could be area or combination of area and perimeter
    max_area = 0

    for contour in contours:
        area = cv2.contourArea(contour)
        perimeter = cv2.arcLength(contour, True)

        # Here we use a combination of area and perimeter to select the contour
        measure = area + perimeter  # You can also try different combinations

        # if measure > max_measure:
        #     max_measure = measure
        #     max_contour = contour

        if area > max_area:
            max_area = area
            max_contour = contour

    contour_img = image.copy()
    cv2.drawContours(contour_img, [max_contour], -1, (0, 255, 0), 2)
    showImage(contour_img)

通用修复方案 #

1. 预处理优化：自适应直方图均衡+双边滤波 #

替换原有高斯模糊，用双边滤波在降噪的同时保留边缘细节；搭配CLAHE自适应直方图均衡，解决光照不均问题，强化文档与背景的对比度：

# 灰度转换后替换为以下预处理步骤
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
# CLAHE自适应直方图均衡
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
gray_eq = clahe.apply(gray)
# 双边滤波：降噪+保边缘
blurred = cv2.bilateralFilter(gray_eq, 9, 75, 75)

2. 边缘检测：动态阈值Canny+形态学开闭运算 #

放弃固定阈值Canny，改用中位数自动计算Canny上下限；形态学操作采用开闭运算组合，避免过度膨胀/腐蚀导致边缘变形：

# 自动计算Canny阈值
v = np.median(blurred)
lower = int(max(0, (1.0 - 0.33) * v))
upper = int(min(255, (1.0 + 0.33) * v))
edges = cv2.Canny(blurred, lower, upper)

# 形态学开闭运算：先去噪点，再补全边缘缺口
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3,3))
opening = cv2.morphologyEx(edges, cv2.MORPH_OPEN, kernel, iterations=1)
closing = cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel, iterations=2)

3. 轮廓筛选：四边形特征+面积占比约束 #

仅靠面积筛选易误选桌面边框，需结合轮廓近似为四边形+面积占图像比例的双重条件，精准定位文档：

contours, _ = cv2.findContours(closing.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
# 按面积降序排序，优先处理大轮廓
contours = sorted(contours, key=cv2.contourArea, reverse=True)

doc_contour = None
image_area = image.shape[0] * image.shape[1]

for cnt in contours:
    perimeter = cv2.arcLength(cnt, True)
    # 轮廓近似，epsilon取周长的2%
    approx = cv2.approxPolyDP(cnt, 0.02 * perimeter, True)
    # 筛选条件：四边形 + 面积占图像10%~90%（排除过小/过大干扰）
    if len(approx) == 4 and 0.1*image_area < cv2.contourArea(cnt) < 0.9*image_area:
        doc_contour = approx
        break

4. 鲁棒性增强：透视变换验证（可选） #

对检测到的四边形做透视变换，验证变换后区域的纹理均匀度，进一步提升多场景下的检测准确性。

完整修复后代码 #

import cv2
import numpy as np

def loadImage(imagePath):
    return cv2.imread(imagePath)

def showImage(image):
    cv2.imshow("Result", image)
    cv2.waitKey(0)
    cv2.destroyAllWindows()

def detectDocEdge(imagePath):
    image = loadImage(imagePath)
    if image is None:
        print("无法加载图像")
        return

    # 1. 预处理：CLAHE+双边滤波
    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
    gray_eq = clahe.apply(gray)
    blurred = cv2.bilateralFilter(gray_eq, 9, 75, 75)

    # 2. 边缘检测：自动阈值Canny+形态学开闭运算
    v = np.median(blurred)
    lower = int(max(0, (1.0 - 0.33) * v))
    upper = int(min(255, (1.0 + 0.33) * v))
    edges = cv2.Canny(blurred, lower, upper)

    kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3,3))
    opening = cv2.morphologyEx(edges, cv2.MORPH_OPEN, kernel, iterations=1)
    closing = cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel, iterations=2)

    # 3. 轮廓筛选：四边形+面积占比
    contours, _ = cv2.findContours(closing.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    contours = sorted(contours, key=cv2.contourArea, reverse=True)

    doc_contour = None
    image_area = image.shape[0] * image.shape[1]

    for cnt in contours:
        perimeter = cv2.arcLength(cnt, True)
        approx = cv2.approxPolyDP(cnt, 0.02 * perimeter, True)
        if len(approx) == 4 and 0.1*image_area < cv2.contourArea(cnt) < 0.9*image_area:
            doc_contour = approx
            break

    # 4. 结果可视化
    if doc_contour is not None:
        result_img = image.copy()
        cv2.drawContours(result_img, [doc_contour], -1, (0,255,0), 3)
        showImage(result_img)
    else:
        print("未检测到文档边缘")
        # 调试用：显示所有轮廓
        contour_img = image.copy()
        cv2.drawContours(contour_img, contours, -1, (0,255,0), 2)
        showImage(contour_img)

# 调用示例
# detectDocEdge("your_image_path.jpg")

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