如何转换灰度图像的直方图以强制执行 highlights/midtones/shadows 的特定比率？

Question

我收集了大量 7 兆像素的灰度图像，我想对它们进行批处理以调整对比度和亮度，以便每张图像包含大约：

50%高光（发光值为200-255的像素点）
30%中间调（发光值55-199的像素点）
20%阴影（发光值为0-54的像素点）

它需要相当高效，因为我只有 1.8GHz 和许多图像。我知道使用 NumPy 可以 PIL/Pillow 比没有 NumPy 更有效地处理图像，但我以前从未使用过它。

Answer 1

不久前我写了一些 numpy 代码来解决这个问题。

有许多可能的方法来转换输入图像的直方图，使正确数量的像素值落在每个区间内。也许最简单的方法是找到每个百分位数对应的当前像素值与所需值之间的差异，然后跨 bin 边缘进行线性插值以找出每个像素值 add/subtract 的多少：

import numpy as np

def hist_norm(x, bin_edges, quantiles, inplace=False):
    """
    Linearly transforms the histogram of an image such that the pixel values
    specified in `bin_edges` are mapped to the corresponding set of `quantiles`

    Arguments:
    -----------
        x: np.ndarray
            Input image; the histogram is computed over the flattened array
        bin_edges: array-like
            Pixel values; must be monotonically increasing
        quantiles: array-like
            Corresponding quantiles between 0 and 1. Must have same length as
            bin_edges, and must be monotonically increasing
        inplace: bool
            If True, x is modified in place (faster/more memory-efficient)

    Returns:
    -----------
        x_normed: np.ndarray
            The normalized array
    """

    bin_edges = np.atleast_1d(bin_edges)
    quantiles = np.atleast_1d(quantiles)

    if bin_edges.shape[0] != quantiles.shape[0]:
        raise ValueError('# bin edges does not match number of quantiles')

    if not inplace:
        x = x.copy()
    oldshape = x.shape
    pix = x.ravel()

    # get the set of unique pixel values, the corresponding indices for each
    # unique value, and the counts for each unique value
    pix_vals, bin_idx, counts = np.unique(pix, return_inverse=True,
                                          return_counts=True)

    # take the cumsum of the counts and normalize by the number of pixels to
    # get the empirical cumulative distribution function (which maps pixel
    # values to quantiles)
    ecdf = np.cumsum(counts).astype(np.float64)
    ecdf /= ecdf[-1]

    # get the current pixel value corresponding to each quantile
    curr_edges = pix_vals[ecdf.searchsorted(quantiles)]

    # how much do we need to add/subtract to map the current values to the
    # desired values for each quantile?
    diff = bin_edges - curr_edges

    # interpolate linearly across the bin edges to get the delta for each pixel
    # value within each bin
    pix_delta = np.interp(pix_vals, curr_edges, diff)

    # add these deltas to the corresponding pixel values
    pix += pix_delta[bin_idx]

    return pix.reshape(oldshape)

例如：

from scipy.misc import lena

bin_edges = 0, 55, 200, 255
quantiles = 0, 0.2, 0.5, 1.0
img = lena()
normed = hist_norm(img, bin_edges, quantiles)

绘图：

from matplotlib import pyplot as plt

def ecdf(x):
    vals, counts = np.unique(x, return_counts=True)
    ecdf = np.cumsum(counts).astype(np.float64)
    ecdf /= ecdf[-1]
    return vals, ecdf

x1, y1 = ecdf(img.ravel())
x2, y2 = ecdf(normed.ravel())

fig = plt.figure()
gs = plt.GridSpec(2, 2)
ax1 = fig.add_subplot(gs[0, 0])
ax2 = fig.add_subplot(gs[0, 1], sharex=ax1, sharey=ax1)
ax3 = fig.add_subplot(gs[1, :])
for aa in (ax1, ax2):
    aa.set_axis_off()

ax1.imshow(img, cmap=plt.cm.gray)
ax1.set_title('Original')
ax2.imshow(normed, cmap=plt.cm.gray)
ax2.set_title('Normalised')

ax3.plot(x1, y1 * 100, lw=2, label='Original')
ax3.plot(x2, y2 * 100, lw=2, label='Normalised')
for xx in bin_edges:
    ax3.axvline(xx, ls='--', c='k')
for yy in quantiles:
    ax3.axhline(yy * 100., ls='--', c='k')
ax3.set_xlim(bin_edges[0], bin_edges[-1])
ax3.set_xlabel('Pixel value')
ax3.set_ylabel('Cumulative %')
ax3.legend(loc=2)

如何转换灰度图像的直方图以强制执行 highlights/midtones/shadows 的特定比率？

How can I transform the histograms of grayscale images to enforce a particular ratio of highlights/midtones/shadows?

python

numpy

image-processing

python-imaging-library

pillow