MAD算法如何自适应调整融合了KMeans的CLAHE的阈值代码python

MAD (Median Absolute Deviation) 是一种鲁棒的统计方法，可以用来自适应调整融合了 KMeans 的 CLAHE 的阈值。

具体步骤如下：

1. 对于每个图像块，使用 KMeans 算法将其分为若干个子块，每个子块的灰度值相近。

2. 对于每个子块，计算其中像素值的 MAD 值，作为该子块的阈值。

3. 对于整张图像，使用 CLAHE 算法进行直方图均衡化，但是使用每个子块的阈值作为对应子块的阈值。

4. 如果整张图像的亮度过于明亮或暗淡，可以根据整张图像的 MAD 值调整阈值。如果 MAD 值较小，则将所有子块的阈值增加一个固定值；如果 MAD 值较大，则将所有子块的阈值减少一个固定值。

代码实现如下：

import numpy as np
from skimage.filters import median
from skimage.filters.rank import median as median_rank
from skimage.exposure import equalize_adapthist
from sklearn.cluster import KMeans

def mad_clahe(image, block_size=32, kmeans_k=4, delta=40):
    # Step 1: Use KMeans to cluster each block of image into k clusters
    kmeans = KMeans(n_clusters=kmeans_k, random_state=0)
    block_shape = (block_size, block_size)
    blocks = view_as_blocks(image, block_shape)
    n_blocks = blocks.shape[0] * blocks.shape[1]
    block_centers = np.zeros((n_blocks, kmeans_k))
    for i in range(blocks.shape[0]):
        for j in range(blocks.shape[1]):
            block = blocks[i, j]
            block_flat = block.reshape((-1, 1))
            kmeans.fit(block_flat)
            block_centers[i * blocks.shape[1] + j] = kmeans.cluster_centers_.ravel()

    # Step 2: Calculate threshold for each block as median absolute deviation
    block_mads = np.median(np.abs(block_centers - np.median(block_centers, axis=1, keepdims=True)), axis=1)

    # Step 3: Apply CLAHE on image with adaptive thresholds
    adaptive_thresholds = block_mads.reshape(blocks.shape[0], blocks.shape[1])
    image_clahe = np.zeros_like(image)
    for i in range(blocks.shape[0]):
        for j in range(blocks.shape[1]):
            block = blocks[i, j]
            block_clahe = equalize_adapthist(block, clip_limit=0.03, kernel_size=None, nbins=256, adaptive_clip_limit=True, clip_limit_range=None, mask=None, nbins_above_high=None, nbins_below_low=None, channel_axis=None)
            image_clahe[i * block_size:(i + 1) * block_size, j * block_size:(j + 1) * block_size] = block_clahe

    # Step 4: Adjust thresholds according to global MAD value
    global_mad = np.median(np.abs(image_clahe - np.median(image_clahe)))
    if global_mad < delta:
        adaptive_thresholds += delta
    else:
        adaptive_thresholds -= delta

    # Final step: Apply median filter on enhanced image
    image_enhanced = median_rank(image_clahe, selem=disk(1))

    return image_enhanced

其中，`image` 是待增强的图像，`block_size` 是每个块的大小，`kmeans_k` 是 KMeans 算法中的聚类数目，`delta` 是调整阈值时的增量。

这段代码使用了 `sklearn.cluster.KMeans` 来进行聚类，使用了 `skimage.filters.rank.median` 来进行局部中值滤波，使用了 `skimage.exposure.equalize_adapthist` 来进行自适应直方图均衡化。