def trainImage(imageNumber):
imageFile = PATH + ('train%d.png' % imageNumber)
mapFile = PATH + ('map%d.txt' % imageNumber)
nums = readMapFile(mapFile)
img = cv2.imread(imageFile)
charBounds = itl_paragraph.parseParagraph(img, returnBounds=True)
# Use this to check character bounding boxes
# for charBound in charBounds:
# if imageNumber == 6:
# cv2.imshow('IMG', charBound)
# cv2.waitKey(0)
if len(charBounds) != len(nums):
print 'ERROR: Image #%d, len(mapFile) = %d, len(detectedChars) = %d' %\
(imageNumber, len(nums), len(charBounds))
exit()
samples = np.empty((0, 400))
# samples = np.empty((0, 10))
for charBound in charBounds:
# Use this to check that the character bounding boxes are accurate
# if imageNumber > 0:
# cv2.imshow('IMG', charBound)
# cv2.waitKey(0)
# itl_char.parseCharacter(charBound)
# Pixel Values as features
charBound = itl_char.getCroppedImage(
itl_char.getBinaryImage(charBound))
charBound = cv2.resize(charBound, CHAR_BOX)
charBound = np.float32(charBound.reshape((1, 400)))
charBound = charBound / np.linalg.norm(charBound)
# Frey-Slate attributes
# charBound = itl_char.getFreySlateAttributes(charBound)
samples = np.append(samples, charBound, 0)
nums = np.array(nums, np.int)
nums = nums.reshape((nums.size, 1))
return (nums, samples)
def trainImage(imageNumber):
imageFile = PATH + ('train%d.png' % imageNumber)
mapFile = PATH + ('map%d.txt' % imageNumber)
nums = readMapFile(mapFile)
img = cv2.imread(imageFile)
charBounds = itl_paragraph.parseParagraph(img, returnBounds=True)
# Use this to check character bounding boxes
# for charBound in charBounds:
# if imageNumber == 6:
# cv2.imshow('IMG', charBound)
# cv2.waitKey(0)
if len(charBounds) != len(nums):
print 'ERROR: Image #%d, len(mapFile) = %d, len(detectedChars) = %d' %\
(imageNumber, len(nums), len(charBounds))
exit()
samples = np.empty((0, 400))
# samples = np.empty((0, 10))
for charBound in charBounds:
# Use this to check that the character bounding boxes are accurate
# if imageNumber > 0:
# cv2.imshow('IMG', charBound)
# cv2.waitKey(0)
# itl_char.parseCharacter(charBound)
# Pixel Values as features
charBound = itl_char.getCroppedImage(itl_char.getBinaryImage(charBound))
charBound = cv2.resize(charBound, CHAR_BOX)
charBound = np.float32(charBound.reshape((1, 400)))
charBound = charBound / np.linalg.norm(charBound)
# Frey-Slate attributes
# charBound = itl_char.getFreySlateAttributes(charBound)
samples = np.append(samples, charBound, 0)
nums = np.array(nums, np.int)
nums = nums.reshape((nums.size, 1))
return (nums, samples)
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
