Mask adjacent blobs when extracting blob images.

pixelocr/image.py

@@ -20,7 +20,7 @@ import numpy as np
 from skimage.io import imread, imsave
 from skimage.color import rgb2gray
 
-from .utils import cached_property
+from .utils import cached_property, pairwise
 
 
 def _is_nonblank(bitmap):
@@ -162,6 +162,11 @@ class Image(object):
         bottom_margin = _get_margin_height(reversed(self.bitmap))
         return self[top_margin:self.height - bottom_margin, :]
 
+    def mask(self, mask):
+        mask3 = np.dstack([mask] * 3)
+        data = np.ma.masked_array(self.data, mask3, fill_value=255).filled()
+        return Image(data, self.x, self.y)
+
     def _iter_lines(self, min_space):
         def iter_lines():
             line_start = None

pixelocr/page.py

@@ -118,8 +118,11 @@ class Word(PageObject):
     @collect_iterable
     def letters(self):
         labels, max_label = ndimage.label(self.image.bitmap, CONNECTIVITY8)
-        obj_indices = ndimage.find_objects(labels, max_label)
-        letter_images = (self.image[obj_index] for obj_index in obj_indices)
+        blob_slices = enumerate(ndimage.find_objects(labels, max_label), 1)
+        letter_images = (
+            self._extract_blob(blob_slice, label, labels)
+            for (label, blob_slice) in blob_slices
+        )
         letters = (
             Letter(image, self.baseline - image.bottom)
             for image in letter_images
@@ -127,6 +130,12 @@ class Word(PageObject):
         sorted_letters = sorted(letters, key=lambda letter: (letter.left, -letter.bottom))
         return iter(sorted_letters)
 
+    def _extract_blob(self, blob_slice, label, labels):
+        image = self.image[blob_slice]
+        mask = labels[blob_slice] != label
+        return image.mask(mask)
+
+
 
 class Letter(PageObject):
     def __init__(self, image, elevation):