Adaptive threshold with masking (two versions)

masked_adaptive_threshold.py

from __future__ import division
import numpy as np

from scipy.ndimage.filters import correlate1d

def masked_adaptive_threshold(image,mask,max_value,size,C):
    '''
    image must already be masked (unmasked areas=0)
    mask is a boolean array
    see http://stackoverflow.com/a/10015315/60982
    '''
    block = np.ones(size, dtype='d')
    
    conv = correlate1d(image.astype(np.uint64), block, axis=0, mode='constant')
    conv = correlate1d(conv, block, axis=1, mode='constant')
    conv -= image # only consider neighbors
    
    number_neighbours = correlate1d(mask.astype(np.uint64), block, axis=0, mode='constant')
    number_neighbours = correlate1d(number_neighbours, block, axis=1, mode='constant')
    number_neighbours -= mask # only count neighbors
    
    mean_conv = conv / number_neighbours

    binary = np.zeros_like(image)
    binary[np.logical_and(image > mean_conv - C, mask)] = max_value
    return binary

import cv2

def masked_adaptive_threshold2(image,mask,max_value,size,C):
    '''
    image must already be masked (unmasked areas=0)
    mask is a boolean array
    
    same as masked_adaptive_threshold but very fast, becomes slightly
    inaccurate on very dark areas
    see http://stackoverflow.com/a/10551103/60982    
    '''
    mask = mask.astype(np.uint8) * 255
    conv = cv2.blur(image, (size, size)).astype(float)
    number_neighbours = cv2.blur(mask, (size, size)).astype(float)   
    image = image-255*(conv/number_neighbours)    
    binary = np.zeros_like(image, dtype=np.uint8)
    binary[np.logical_and(image > -C, mask)] = max_value   
    return binary