How Can I Transform The Histograms Of Grayscale Images To Enforce A Particular Ratio Of Highlights/midtones/shadows?

I have a large collection of 7 mega pixel grayscale images and I want batch process them to adjust the contrast and brightness so that each image contains about: 50% highlights (p

Solution 1:

A while ago I wrote some numpy code to solve this exact problem.

There are many possible ways to transform the histogram of the input image such that the correct number of pixel values fall within each bin. Perhaps the simplest is to find the difference between the current pixel values corresponding to each percentile and the desired value, then linearly interpolate across the bin edges to find out how much to add/subtract from each pixel value:

import numpy as np

defhist_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')

    ifnot 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)

For example:

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)

Plotting:

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)