如何在Python中预处理亮度与对比度以精准识别图像差异？

针对你遇到的「同图调整亮度/对比度后被SSIM误判为有差异」的问题，以下是几个基于skimage的预处理方案，能有效匹配图像的亮度、对比度甚至颜色分布，提升差异识别的精准度：

1. 直方图匹配（最直接解决亮度/对比度差异） #

直方图匹配能将两张图像的像素值分布强制对齐，完全消除因亮度、对比度调整带来的全局分布差异，是解决这类问题的首选方案。使用skimage.exposure.match_histograms实现：

修改你的预处理代码部分：

from skimage.exposure import match_histograms

def get_structural_simlarity(first_image, second_image):
    print("[Console] Calculating differences")
    # Pre-process Image
    first_gray = convert_to_gray(first_image)
    second_gray = convert_to_gray(second_image)
    
    # 关键：将第二张图的直方图匹配到第一张图的分布
    second_gray_matched = match_histograms(second_gray, first_gray)
    
    # Compare
    (score, diff_img) = structural_similarity(first_gray, second_gray_matched, full=True)
    # Convert return format to cv2 readable
    diff_img = convert_to_cv2_format(diff_img)
    print("[Console] Similarity score of {:.4f}%".format(score * 100))
    return diff_img

2. 全局归一化+标准化（应对极端亮度差异） #

如果图像存在全局过亮/过暗的情况，可以先做强度归一化，再用z-score标准化，把两张图的像素值映射到相同的均值和标准差范围：

from skimage.exposure import rescale_intensity
from skimage.util import img_as_float

def normalize_image(image):
    # 转为浮点型并归一化到[0,1]区间
    img_float = img_as_float(image)
    img_normalized = rescale_intensity(img_float, in_range='image', out_range=(0, 1))
    # z-score标准化，统一均值和标准差
    mean = img_normalized.mean()
    std = img_normalized.std()
    return (img_normalized - mean) / std

def get_structural_simlarity(first_image, second_image):
    print("[Console] Calculating differences")
    # Pre-process Image
    first_gray = convert_to_gray(first_image)
    second_gray = convert_to_gray(second_image)
    
    # 对两张图分别做归一化+标准化
    first_gray_norm = normalize_image(first_gray)
    second_gray_norm = normalize_image(second_gray)
    
    # Compare
    (score, diff_img) = structural_similarity(first_gray_norm, second_gray_norm, full=True)
    diff_img = convert_to_cv2_format(diff_img)
    print("[Console] Similarity score of {:.4f}%".format(score * 100))
    return diff_img

3. 彩色图像专用：Lab颜色空间分离处理 #

如果你的输入是彩色图像，直接转灰度会丢失颜色信息，建议转成Lab颜色空间，单独对亮度通道（L）做直方图匹配，对颜色通道（a/b）做归一化，这样同时覆盖亮度、对比度和颜色差异：

from skimage.color import rgb2lab, lab2rgb
from skimage.exposure import match_histograms

def preprocess_color_image(src_img, target_img):
    # 转Lab颜色空间，分离亮度与颜色通道
    src_lab = rgb2lab(src_img)
    target_lab = rgb2lab(target_img)
    
    # 匹配L通道（亮度）的直方图
    target_lab[..., 0] = match_histograms(target_lab[..., 0], src_lab[..., 0])
    # 归一化a/b颜色通道，对齐均值和标准差
    for channel in [1, 2]:
        src_mean = src_lab[..., channel].mean()
        src_std = src_lab[..., channel].std()
        target_channel = target_lab[..., channel]
        target_lab[..., channel] = (target_channel - target_channel.mean()) / target_channel.std() * src_std + src_mean
    
    # 转回RGB并转灰度用于后续对比
    target_matched_rgb = lab2rgb(target_lab)
    return convert_to_gray(target_matched_rgb)

def get_structural_simlarity(first_image, second_image):
    print("[Console] Calculating differences")
    # Pre-process Image
    first_gray = convert_to_gray(first_image)
    # 对第二张图做彩色预处理后转灰度
    second_gray_matched = preprocess_color_image(first_image, second_image)
    
    # Compare
    (score, diff_img) = structural_similarity(first_gray, second_gray_matched, full=True)
    diff_img = convert_to_cv2_format(diff_img)
    print("[Console] Similarity score of {:.4f}%".format(score * 100))
    return diff_img

为什么equalize_adapthist没效果？ #

equalize_adapthist是自适应局部直方图均衡，核心作用是增强图像的局部对比度，但它不会对齐两张图的全局亮度/对比度分布。两张图各自做自适应均衡后，原本的亮度差异可能依然存在，甚至会引入新的局部差异，因此无法解决你的问题。

内容的提问来源于stack exchange，提问作者Anh Phan