def parseWord(img, returnBounds=False):
img1 = cv2.cvtColor(img, cv.CV_BGR2GRAY)
if returnBounds:
# TRAINING parameters, with nicely spaced characters
unused, img1 = cv2.threshold(img1, 200, 255, cv.CV_THRESH_BINARY_INV)
img1 = cv2.GaussianBlur(img1, (5, 5), 0)
img1 = cv2.dilate(img1, (2, 2), 1)
element = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
img1 = cv2.morphologyEx(img1, cv.CV_MOP_CLOSE, element)
# cv2.imshow('MORPH', img1)
# cv2.waitKey(0)
else:
img1 = cv2.GaussianBlur(img1, (1, 1), 0)
# img1 = cv2.Laplacian(img1, cv2.CV_8U)
# img1 = cv2.equalizeHist(img1)
unused, img1 = cv2.threshold(img1, 200, 255, cv.CV_THRESH_BINARY_INV)
# img1 = cv2.dilate(img1, (1, 1), 1)
if DEBUG:
cv2.imwrite(IMAGE_NAME + '-thresh.' + EXTENSION, img1)
# Blur in the horizontal direction to get lines
element = cv2.getStructuringElement(cv2.MORPH_RECT, (1, 1))
img1 = cv2.morphologyEx(img1, cv.CV_MOP_CLOSE, element)
if DEBUG:
cv2.imwrite(IMAGE_NAME + '-morph.' + EXTENSION, img1)
# Use RETR_EXTERNAL to remove boxes that are completely contained by the word
contours, hierarchy = cv2.findContours(img1, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
boundRects = []
for i in xrange(len(contours)):
contourPoly = cv2.approxPolyDP(contours[i], 3, True)
boundRect = cv2.boundingRect(contourPoly)
boundRects.append(boundRect)
# Pad the character bounding boxes
height = img.shape[0]
width = img.shape[1]
boundRects = sorted(boundRects, key=lambda x: x[0])
pad = 1
adjustedRects = []
minX = width
maxX = 0
for rect in boundRects:
adjustedRect = (rect[0] - pad, rect[1] - pad, rect[2] + pad * 2, rect[3] + pad * 2)
adjustedRects.append(adjustedRect)
minX = min(minX, adjustedRect[0])
maxX = max(maxX, adjustedRect[0] + adjustedRect[2])
if DEBUG:
# for rect in adjustedRects:
# cv2.rectangle(img, (rect[0], rect[1]), (rect[0] + rect[2], rect[1] + rect[3]), (0, 255, 0))
cv2.imwrite(IMAGE_NAME + '-bounds.' + EXTENSION, img)
print '%d char bounding boxes initially found in %s' % (len(adjustedRects), IMAGE_FILE)
# for rect in adjustedRects:
# cv2.rectangle(img, (rect[0], rect[1]), (rect[0] + rect[2], rect[1] + rect[3]), (0, 255, 0))
# cv2.imshow('IMG', img)
# cv2.waitKey(0)
# Extract characters from word
# chars = []
# for rect in adjustedRects:
# [x, y, w, h] = rect
# char = img[0:height, x:(x+w)]
# chars.append(char)
latex = []
img = itl_char.getCroppedImage(itl_char.getBinaryImage(img))
width = img.shape[1]
np.set_printoptions(threshold=np.inf, linewidth=np.inf)
columnSums = np.sum(img, axis=0)
currentX = 0
MIN_WIDTH = 2
MIN_CHAR_WIDTH = 4
MAX_CHAR_WIDTH = 15
THRESHOLD = .8
while width - currentX >= MIN_WIDTH:
if columnSums[currentX] == 0 or columnSums[currentX + 1] == 0:
currentX += 1
continue
start = MIN_CHAR_WIDTH
end = MAX_CHAR_WIDTH
if (currentX + 2 < width and columnSums[currentX + 2] == 0) or\
(currentX + 3 < width and columnSums[currentX + 3] == 0):
start = MIN_WIDTH
end = MIN_CHAR_WIDTH + 1
scores = [False] * (MAX_CHAR_WIDTH + 1)
values = [None] * (MAX_CHAR_WIDTH + 1)
brokeSpace = False
goodCharScore = False
for w in xrange(start, end):
rightX = currentX + w
charSlice = img[0:height, currentX:rightX]
# print charSlice
val, score = itl_char.parseCharacter(charSlice, getScore=True, isBinary=True)
scores[w] = score
if score > THRESHOLD and w >= MIN_CHAR_WIDTH:
goodCharScore = True
values[w] = val
if rightX >= width or columnSums[rightX] == 0:
brokeSpace = True
break
maxRunLength = 0
maxRunBestScore = 0.0
maxRunBestWidth = 0
currentRunLength = 0
currentRunBestScore = 0.0
currentRunBestWidth = 0
if brokeSpace:
maxRunBestWidth = w
maxRunBestScore = scores[w]
bestVal = values[w]
else:
for w in xrange(start, end + 1):
if scores[w] < THRESHOLD or values[w][0] in '.,:;':
currentRunLength = 0
currentRunBestScore = 0.0
continue
currentRunLength += 1
if scores[w] > currentRunBestScore:
currentRunBestScore = scores[w]
currentRunBestWidth = w
if currentRunLength > maxRunLength or\
(currentRunLength == maxRunLength and currentRunBestScore > maxRunBestScore):
maxRunLength = currentRunLength
maxRunBestScore = currentRunBestScore
maxRunBestWidth = currentRunBestWidth
if maxRunBestWidth == 0:
break
bestVal = values[maxRunBestWidth]
# print '*****', maxRunBestWidth, maxRunBestScore, bestVal
# Add to chars array
currentX += maxRunBestWidth
latex.append(bestVal)
if returnBounds:
return chars
# latex = []
# for i in xrange(len(chars)):
# char = chars[i]
# charLatex = itl_char.parseCharacter(char)
# latex.append(charLatex)
# if DEBUG:
# cv2.imshow('%d' % i, char)
# cv2.waitKey(0)
return latex
def parseEquation(img):
img_gray = cv2.cvtColor(img, cv.CV_BGR2GRAY)
#img_inv = 255 - img_gray
# img_lap = cv2.Laplacian(img_gray, cv2.CV_8U)
# if DEBUG:
# cv2.imwrite(IMAGE_NAME + '-lap.' + EXTENSION, img_lap)
unused, img_threshold = cv2.threshold(img_gray, 220, 255,
cv.CV_THRESH_BINARY_INV)
if DEBUG:
cv2.imwrite(IMAGE_NAME + '-thresh.' + EXTENSION, img_threshold)
# Blur in the horizontal direction to get lines
morph_size = (0, 0)
# element = cv2.getStructuringElement(cv2.MORPH_RECT, morph_size)
# morphed = cv2.morphologyEx(img_threshold, cv.CV_MOP_CLOSE, element)
# if DEBUG:
# cv2.imwrite(IMAGE_NAME + '-morph.' + EXTENSION, morphed)
# Use RETR_EXTERNAL to remove boxes that are completely contained by the word
contours, hierarchy = cv2.findContours(img_threshold, cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_NONE)
boundRects = []
for i in xrange(len(contours)):
contourPoly = cv2.approxPolyDP(contours[i], 0.25, True)
boundRect = cv2.boundingRect(contourPoly)
if boundRect[2] * boundRect[3] > 1:
boundRects.append(
(boundRect[0] - morph_size[0], boundRect[1] - morph_size[1],
boundRect[2] + morph_size[0], boundRect[3]))
# # Filter bounding rectangles that are not an entire line
# # Take the maximum height among all bounding boxes
# # Remove those boxes that have height less than 25% of the maximum
# maxHeight = -1
# for rect in boundRects:
# maxHeight = max(rect[3], maxHeight)
# heightThresh = .25 * maxHeight
# boundRects = [rect for rect in boundRects if rect[3] > heightThresh]
# if DEBUG:
# for rect in boundRects:
# cv2.rectangle(img, (rect[0], rect[1]), (rect[0] + rect[2], rect[1] + rect[3]), (0, 255, 0))
# cv2.imwrite(IMAGE_NAME + '-bounds.' + EXTENSION, img)
# print '%d words found in %s' % (len(boundRects), IMAGE_FILE)
sortedRects = sorted(boundRects, key=lambda x: (x[0], x[1]))
# words = []
# for rect in sortedRects:
# [x, y, w, h] = rect
# word = img[y:(y+h), x:(x+w)]
# words.append(word)
# for word in words:
# # TODO do something here
# if DEBUG:
# cv2.imshow('Word', word)
# cv2.waitKey(0)
# keys = ['n', '-', '-', '\\sum', '\\infty', '-', '\\infty', '|', '\\langle',
# 'f', ',', '-', 'e', '\\sqrt', 'i', 'n', '2', '\\pi', 'x', '\\rangle',
# '|', '2', '-', '-', '|', '|', 'f', '|', '|', '2'
# ]
keys = []
for j in xrange(len(sortedRects)):
x, y, w, h = sortedRects[j]
img_bb = img[y:y + h, x:x + w]
key = itl_char.parseCharacter(img_bb)
keys.append(translateKey(key))
L = []
for j in xrange(len(keys)):
L.append(buildSymbol(keys[j], *(sortedRects[j])))
L = handleSpecialCases(L)
tree, baseline = findSymbolTree(L)
for i in xrange(len(L)):
c = L[i].centroid()
rounded = (int(round(c[0])), int(round(c[1])))
if i in baseline:
cv2.circle(img, rounded, 2, (255, 0, 0), 1)
else:
cv2.circle(img, rounded, 2, (0, 0, 255), 1)
for i in xrange(len(L)):
for j in tree[i]:
c1 = L[i].centroid()
c2 = L[j].centroid()
r1 = (int(round(c1[0])), int(round(c1[1])))
r2 = (int(round(c2[0])), int(round(c2[1])))
if i in baseline and j in baseline:
cv2.line(img, r1, r2, (255, 0, 0))
else:
cv2.line(img, r1, r2, (0, 0, 255))
cv2.imwrite(IMAGE_NAME + '-mst.' + EXTENSION, img)
node = buildLaTeXTree(L)
node_str = node.str()
print 'latex string:', node_str
return node_str
def parseEquation(img):
img_gray = cv2.cvtColor(img, cv.CV_BGR2GRAY)
#img_inv = 255 - img_gray
# img_lap = cv2.Laplacian(img_gray, cv2.CV_8U)
# if DEBUG:
# cv2.imwrite(IMAGE_NAME + '-lap.' + EXTENSION, img_lap)
unused, img_threshold = cv2.threshold(img_gray, 220, 255, cv.CV_THRESH_BINARY_INV)
if DEBUG:
cv2.imwrite(IMAGE_NAME + '-thresh.' + EXTENSION, img_threshold)
# Blur in the horizontal direction to get lines
morph_size = (0, 0)
# element = cv2.getStructuringElement(cv2.MORPH_RECT, morph_size)
# morphed = cv2.morphologyEx(img_threshold, cv.CV_MOP_CLOSE, element)
# if DEBUG:
# cv2.imwrite(IMAGE_NAME + '-morph.' + EXTENSION, morphed)
# Use RETR_EXTERNAL to remove boxes that are completely contained by the word
contours, hierarchy = cv2.findContours(img_threshold, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
boundRects = []
for i in xrange(len(contours)):
contourPoly = cv2.approxPolyDP(contours[i], 0.25, True)
boundRect = cv2.boundingRect(contourPoly)
if boundRect[2] * boundRect[3] > 1:
boundRects.append((boundRect[0]-morph_size[0], boundRect[1]-morph_size[1], boundRect[2]+ morph_size[0], boundRect[3]))
# # Filter bounding rectangles that are not an entire line
# # Take the maximum height among all bounding boxes
# # Remove those boxes that have height less than 25% of the maximum
# maxHeight = -1
# for rect in boundRects:
# maxHeight = max(rect[3], maxHeight)
# heightThresh = .25 * maxHeight
# boundRects = [rect for rect in boundRects if rect[3] > heightThresh]
# if DEBUG:
# for rect in boundRects:
# cv2.rectangle(img, (rect[0], rect[1]), (rect[0] + rect[2], rect[1] + rect[3]), (0, 255, 0))
# cv2.imwrite(IMAGE_NAME + '-bounds.' + EXTENSION, img)
# print '%d words found in %s' % (len(boundRects), IMAGE_FILE)
sortedRects = sorted(boundRects, key=lambda x:(x[0], x[1]))
# words = []
# for rect in sortedRects:
# [x, y, w, h] = rect
# word = img[y:(y+h), x:(x+w)]
# words.append(word)
# for word in words:
# # TODO do something here
# if DEBUG:
# cv2.imshow('Word', word)
# cv2.waitKey(0)
# keys = ['n', '-', '-', '\\sum', '\\infty', '-', '\\infty', '|', '\\langle',
# 'f', ',', '-', 'e', '\\sqrt', 'i', 'n', '2', '\\pi', 'x', '\\rangle',
# '|', '2', '-', '-', '|', '|', 'f', '|', '|', '2'
# ]
keys = []
for j in xrange(len(sortedRects)):
x, y, w, h = sortedRects[j]
img_bb = img[y:y+h, x:x+w]
key = itl_char.parseCharacter(img_bb)
keys.append(translateKey(key))
L = []
for j in xrange(len(keys)):
L.append(buildSymbol(keys[j], *(sortedRects[j])))
L = handleSpecialCases(L)
tree, baseline = findSymbolTree(L)
for i in xrange(len(L)):
c = L[i].centroid()
rounded = (int(round(c[0])), int(round(c[1])))
if i in baseline:
cv2.circle(img, rounded, 2, (255, 0, 0), 1)
else:
cv2.circle(img, rounded, 2, (0, 0, 255), 1)
for i in xrange(len(L)):
for j in tree[i]:
c1 = L[i].centroid()
c2 = L[j].centroid()
r1 = (int(round(c1[0])), int(round(c1[1])))
r2 = (int(round(c2[0])), int(round(c2[1])))
if i in baseline and j in baseline:
cv2.line(img, r1, r2, (255, 0, 0))
else:
cv2.line(img, r1, r2, (0, 0, 255))
cv2.imwrite(IMAGE_NAME + '-mst.' + EXTENSION, img)
node = buildLaTeXTree(L)
node_str = node.str()
print 'latex string:', node_str
return node_str
