noio · March 6, 2015 20:00 · Mar 6, 2015
diff --git a/gistfile1.py b/gistfile1.py
@@ -0,0 +1,95 @@
+#!/usr/bin/env python
+
+import sys, os
+import numpy as np
+from matplotlib import cm
+from scipy import ndimage, misc, spatial
+
+OKGREEN = '\033[92m'
+WARNING = '\033[93m'
+ENDC = '\033[0m'
+BOLD = '\033[1m'
+
+filename = sys.argv[1]
+fileroot = os.path.splitext(filename)[0]
+
+image = misc.imread(filename)
+
+if len(sys.argv) > 2:
+	hx = sys.argv[2]
+	channels = [hx[0:2], hx[2:4], hx[4:6], "FF"]
+	crack_color = [int(c, 16) for c in channels]
+else:
+	crack_color = image[0,0]
+
+print "Cracks with color: %s" % (crack_color)
+
+cracks = np.all(image == crack_color, axis=-1)
+transparent = image[:,:,3] == 0
+
+regions = 1 - np.logical_or(cracks,	transparent)
+
+labels, num_labels = ndimage.label(regions)
+
+
+# Structuring element that expands to bottom right
+# when used for dilation
+strel = np.array([[0,0,0],
+				  [0,1,1],
+				  [1,1,1]])
+
+
+# We're going to assign all the cracks to the closest region
+# in a way so that each crack pixel belongs to exactly one region
+
+# Create a list of coordinates
+positions = np.indices(labels.shape).transpose([1,2,0])
+
+# Filter that list by the pixels that belong to a region
+positions = positions[labels > 0]
+
+# A KDTree is super overkill but it does save on some code here.
+# Feed the KDTree the coordinates of all non-zero label positions
+kdtree = spatial.KDTree(positions)
+
+# Get positions of crack pixels
+crack_positions = np.transpose(np.nonzero(cracks))
+
+# Offset the cracks a tiny bit to bias the cracks to stick to
+# the closest pixel on the top-left
+crack_positions_offset = crack_positions.astype(float) + [[-0.1, -0.4]]
+
+dists, idxs = kdtree.query(crack_positions_offset)
+
+closest_region_position = positions[idxs]
+closest_region_label = labels[closest_region_position[:,0], closest_region_position[:,1]]
+
+# Assign the regions to the label matrix
+labels[ crack_positions[:,0], crack_positions[:,1] ] = closest_region_label
+
+# Save image that are transparent except for the region
+for label in range(1,num_labels+1):
+	to_delete = (labels != label)
+	isolated = np.copy(image)
+	# Set all other pixels to transparent
+	isolated[to_delete] = [0,0,0,0]
+	# Compute a centroid
+	cy, cx = np.average(np.transpose(np.nonzero(labels == label)), axis=0).astype(int)
+	cy = round(cy / 4) * 4
+	cx = round(cx / 4) * 4
+	# Save the image using the centroid as name
+	# We do this so that the names of unaffected pieces
+	# stay the same if other pieces are merged or split
+	savepath = '%s_s%03d_%03d.png' % (fileroot, cx, cy)
+	if os.path.exists(savepath):
+		status = WARNING + BOLD + "Overwriting" + ENDC
+	else:
+		status = OKGREEN + "New" + ENDC
+	print "[%d/%d] %s %s" % (label, num_labels, status, os.path.basename(savepath))
+	misc.imsave(savepath, isolated)
+
+# Save an image of the generated regions
+colormap = cm.get_cmap("Set3")
+colorlabels = colormap(labels / float(np.max(labels)))
+colorlabels[:,:,-1] = image[:,:,-1] / 255.0
+misc.imsave('%s_labels.png' % (fileroot), colorlabels)
No results found