Fix sloppy merge

fourcell/normalize.py

@@ -295,9 +295,6 @@ def create_blank (w, h, rgb_color = (255, 255, 255)) :
 	blank[:] = color
 	return blank
 
-def place (bottom, top, x, y) :
-	return bottom.paste(top, (x, y))
-
 def get_mean_rect (holePunches) :
 	left = 0
 	right = 0
@@ -308,7 +305,7 @@ def get_mean_rect (holePunches) :
 			left += float(hp['x'])
 			top += float(hp['y'])
 		elif hp['order'] == 2 :
-			right += float(hp['y'])
+			right += float(hp['x'])
 			top += float(hp['y'])
 		elif hp['order'] == 3 :
 			left += float(hp['x'])
@@ -316,9 +313,9 @@ def get_mean_rect (holePunches) :
 		elif hp['order'] == 5 :
 			right += float(hp['x'])
 			bottom += float(hp['y'])
-	x = round((right / 2.0) - (left / 2.0))
+	w = round((right / 2.0) - (left / 2.0))
-	y = round((bottom / 2.0) - (top / 2.0))
+	h = round((bottom / 2.0) - (top / 2.0))
-	return (x, y)
+	return (w, h)
 
 
 def center_within (larger, smaller) :
@@ -328,7 +325,19 @@ def center_within (larger, smaller) :
 	h2 = smaller[1]
 	x = ((w1 - w2) / 2)
 	y = ((h1 - h2) / 2)
-	return (x, y)
+	return (int(x), int(y))
+
+# If we consider (0,0) as top left corner of image called 
+# im with left-to-right as x direction and top-to-bottom 
+# as y direction. and we have (x1,y1) as the top-left vertex 
+# and (x2,y2) as the bottom-right vertex of a rectangle 
+# region within that image, then:
+#
+# roi = im[y1:y2, x1:x2]
+
+def crop (img, xoffset, yoffset, w, h) :
+	#crop_img = img[y:y+h, x:x+w].copy()
+	return im[yoffset:yoffset+w, xoffset:xoffset+w].copy()
 
 if len(sys.argv) < 2:
 	print('Please provide path of scan to normalize')
@@ -358,6 +367,7 @@ normalHeight = round(float(width) / pageRatio)
 
 holePunches = find_hole_punches(img)
 rotated = correct_rotation(img, original, holePunches)
+rotatedHeight, rotatedWidth = rotated.shape[:2]
 
 holePunches = find_hole_punches(rotated)
 blank = create_blank(width, normalHeight)
@@ -378,11 +388,13 @@ print(f'Mean rectangle: {meanRect[0]},{meanRect[1]}')
 # offset is the position within the new normal image
 # the top left hole punch should be centered to
 offset = center_within((width, normalHeight), meanRect)
-print(f'Offset: {offset[0]},{offset[1]}')
+print(f'Offset : {offset[0]},{offset[1]}')
+print(f'Topleft: {tl["x"]},{tl["y"]}')
+print(f'Rotated: {rotatedWidth},{rotatedHeight}')
+print(f'Blank  : {width},{normalHeight}')
 
 
 
-display(img)
 #normal = place(blank, rotated, 0, 0)
 
-#display(normal)
+display(blank)

fourcell/notes/brute_force_scale_invariant_template_matching.py

@@ -0,0 +1,81 @@
+import math
+import numpy
+import scipy.ndimage
+import imagecodecs
+import imreg
+from matplotlib import pyplot, patches
+
+
+def brute_force_scale_invariant_template_matching(
+    template,  # grayscale image
+    search,  # scaled and cropped grayscale image
+    zooms=(1.0, 0.5, 0.25),  # sequence of zoom factors to try
+    size=None,  # power-of-two size of square sliding window
+    delta=None,  # advance of sliding windows. default: half window size
+    min_overlap=0.25,  # minimum overlap of search with window
+    max_diff=0.05,  # max average of search - window differences in overlap
+    max_angle=0.5,  # no rotation
+):
+    """Return yoffset, xoffset, and scale of first match of search in template.
+
+    Iterate over scaled versions of the template image in overlapping sliding
+    windows and run FFT-based algorithm for translation, rotation and
+    scale-invariant image registration until a match of the search image is
+    found in the sliding window.
+
+    """
+    if size is None:
+        size = int(pow(2, int(math.log(min(search.shape), 2))))
+    if delta is None:
+        delta = size // 2
+    search = search[:size, :size]
+    for zoom in zooms:
+        windows = numpy.lib.stride_tricks.sliding_window_view(
+            scipy.ndimage.zoom(template, zoom), search.shape
+        )[::delta, ::delta]
+        for i in range(windows.shape[0]):
+            for j in range(windows.shape[1]):
+                print('.', end='')
+                window = windows[i, j]
+                im2, scale, angle, (t0, t1) = imreg.similarity(window, search)
+                diff = numpy.abs(im2 - window)[im2 != 0]
+                if (
+                    abs(angle) < max_angle
+                    and diff.size / window.size > min_overlap
+                    and numpy.mean(diff) < max_diff
+                ):
+                    return (
+                        (i * delta - t0) / zoom,
+                        (j * delta - t1) / zoom,
+                        1 / scale / zoom,
+                    )
+    raise ValueError('no match of search image found in template')
+
+
+def rgb2gray(rgb, scale=None):
+    """Return float grayscale image from RGB24 or RGB48 image."""
+    scale = numpy.iinfo(rgb.dtype).max if scale is None else scale
+    scale = numpy.array([[[0.299, 0.587, 0.114]]], numpy.float32) / scale
+    return numpy.sum(rgb * scale, axis=-1)
+
+
+template = imagecodecs.imread('cw1_IMG_9037.jpg')
+search = imagecodecs.imread('cw1_p1_9037_kzw.jpg')
+
+yoffset, xoffset, scale = brute_force_scale_invariant_template_matching(
+    rgb2gray(template), rgb2gray(search), zooms=(0.5,)
+)
+print(yoffset, xoffset, scale)
+
+figure, ax = pyplot.subplots()
+ax.imshow(template)
+rect = patches.Rectangle(
+    (xoffset, yoffset),
+    scale * search.shape[1],
+    scale * search.shape[0],
+    linewidth=1,
+    edgecolor='r',
+    facecolor='none',
+)
+ax.add_patch(rect)
+pyplot.show()

fourcell/notes/sift.py

@@ -0,0 +1,33 @@
+# Initiate SIFT detector
+sift = cv2.SIFT_create()
+# find the keypoints and descriptors with SIFT
+# Here img1 and img2 are grayscale images
+kp1, des1 = sift.detectAndCompute(img1,None)
+kp2, des2 = sift.detectAndCompute(img2,None)
+
+# FLANN parameters
+# I literally copy-pasted the defaults
+FLANN_INDEX_KDTREE = 1
+index_params = dict(algorithm = FLANN_INDEX_KDTREE, trees = 5)
+search_params = dict(checks=50)   # or pass empty dictionary
+# do the matching
+flann = cv2.FlannBasedMatcher(index_params,search_params)
+matches = flann.knnMatch(des1,des2,k=2)
+
+## OPTIONAL - Show matches ##
+
+# Need to draw only good matches, so create a mask
+matchesMask = [[0,0] for i in range(len(matches))]
+# ratio test as per Lowe's paper <- this is a criterion for matches selection
+for i,(m,n) in enumerate(matches):
+    if m.distance < 0.7*n.distance:
+        matchesMask[i]=[1,0]
+draw_params = dict(matchColor = (0,255,0),
+                   singlePointColor = (255,0,0),
+                   matchesMask = matchesMask,
+                   flags = cv2.DrawMatchesFlags_DEFAULT)
+
+img3 = cv2.drawMatchesKnn(img1,kp1,img2,kp2,matches,None, **draw_params)
+f, ax = plt.subplots(1, figsize=(15,15))
+ax.imshow(img3)
+plt.show()