def create(dataset, work_dir):
# Find all the pages in the dataset
img_dir = Path(Path.cwd().ancestor(1), 'data/hwr_data/pages', dataset)
ann_dir = Path(Path.cwd().ancestor(1), 'data/charannotations')
images = img_dir.listdir('*.jpg')
annotations = ann_dir.listdir(dataset + '*.words')
files = merge(images, annotations)
# Create character segmentations
stats = {}
for f in files:
# Preprocess
logging.info("Preprocessing %s", str(f[0]))
pagesPathFolder = Path(work_dir, 'pages')
pagesPathFolder.mkdir()
pagePath = Path(pagesPathFolder, f[0].stem + '.ppm')
img = cv2.imread(f[0], cv2.IMREAD_GRAYSCALE)
img = preprocess(img)
cv2.imwrite(pagePath, img)
# Segment
segmentPathFolder = Path(work_dir, 'segments')
segmentPathFolder.mkdir()
e = ET.parse(f[1]).getroot()
logging.info("Segmenting %s", str(f[0]))
segment(img, e, segmentPathFolder, stats)
print_statistics(stats, dataset)
def main(state="external",
img=None,
words_file_name=None,
minCutWindow=20,
maxCutWindow=80,
globalLexicon=True):
# Directories
sentenceDir = Path('tmp/sentences/')
wordDir = Path('tmp/words/')
# Logging
logging.basicConfig(format='%(asctime)s %(levelname)-8s: %(message)s',
level=logging.DEBUG,
datefmt='%m/%d/%Y %I:%M:%S %p')
# Check commandline parameters
# if state is external
if state == "external":
if len(sys.argv) != 4:
print "Usage: python %s image.ppm input.words out.words" % sys.argv[
0]
sys.exit(1)
# Read input files
img = cv2.imread(sys.argv[1], 0)
words_file_name = sys.argv[2]
# Preprocess image
img = prep.preprocess(img)
# Get the sorted unique list of class labels
featureDir = 'tmp/features/'
trainDir = featureDir + 'train/'
trainLabels = np.load(trainDir + 'labels.npy')
classes = sorted(set(trainLabels))
# Get lexicon information
lexicon = create_lexicon.create()
lexicon_means_stds = create_lexicon_means_stds.create()
if globalLexicon == True:
stateProbs = create_probTables.create_stateProbs(lexicon)
transProbs = create_probTables.create_transProbs(lexicon)
# Recognition accuracy names
correct = 0
false = 0
inLex = 0
avgReduction = 0
# Recognize the words
xml = ET.parse(words_file_name).getroot()
recog = Recognizer(sentenceDir, wordDir, xml, img)
for word, word_img in recog.next_word():
required_word = word.get('text')
logging.info("Word:\t\t%s" % required_word)
cuts = recog.find_cuts(word_img)
if cuts is not None:
reduced_lexicon, reduction = reduce_lexicon(
cuts, word_img, lexicon, lexicon_means_stds)
avgReduction += reduction
if globalLexicon == False:
stateProbs = create_probTables.create_stateProbs(
reduced_lexicon)
transProbs = create_probTables.create_transProbs(
reduced_lexicon)
cuts.insert(0, 0) # Also create a window at the start of the word
estimate = recog.recursiveRecognize(word_img, cuts,
reduced_lexicon, stateProbs,
transProbs, classes,
minCutWindow, maxCutWindow)
logging.info("Estimate:\t%s" % estimate)
if required_word in reduced_lexicon:
inLex += 1
if required_word == estimate:
correct += 1
else:
false += 1
accuracy = float(correct) / (correct + false) * 100
logging.info("Correct: %s\tFalse: %s\tAccuracy: %s\n" %
(correct, false, accuracy))
else:
continue
accuracy = float(correct) / (correct + false) * 100
totalInLex = float(inLex) / (correct + false) * 100
avgReduction = avgReduction / float(correct + false)
if state == "external":
logging.info(
"Complete results for file: %s, Correct: %s\tFalse: %s \tAccuracy: %s"
% (sys.argv[1], correct, false, accuracy))
ET.ElementTree(recog.words).write(sys.argv[3])
else:
return (correct, false)
logging.info(
"Complete results for file: Correct: %s\tFalse: %s \tAccuracy: %s"
% (correct, false, accuracy))
if __name__ == '__main__':
if len(sys.argv) != 3:
print('Usage: python {} <image> <words>'.format(sys.argv[0]))
sys.exit(1)
img_file = sys.argv[1]
word_file = sys.argv[2]
logger.info('Loading pretrained models...')
with open('tmp/svm.pickle', 'r') as f:
svm = pickle.load(f)
logger.info('Loading image and words file')
img = cv2.imread(img_file, cv2.IMREAD_GRAYSCALE)
img = prep.preprocess(img)
xml = ET.parse(word_file).getroot()
out_dir.mkdir()
for f in out_dir.walk('*'):
f.remove()
logger.info('Starting to create a split file')
for sentence in xml:
for word in sentence:
text = word.get('text')
print(text)
#if '@' in text or len(text) < 6:
if text != 'buton':
continue # Skip short words
def main(
state="external", img=None, words_file_name=None,
minCutWindow = 20, maxCutWindow = 80, globalLexicon=True
):
# Directories
sentenceDir = Path('tmp/sentences/')
wordDir = Path('tmp/words/')
# Logging
logging.basicConfig(format='%(asctime)s %(levelname)-8s: %(message)s', level=logging.DEBUG, datefmt='%m/%d/%Y %I:%M:%S %p')
# Check commandline parameters
# if state is external
if state == "external":
if len(sys.argv) != 4:
print "Usage: python %s image.ppm input.words out.words" % sys.argv[0]
sys.exit(1)
# Read input files
img = cv2.imread(sys.argv[1], 0)
words_file_name = sys.argv[2]
# Preprocess image
img = prep.preprocess(img)
# Get the sorted unique list of class labels
featureDir = 'tmp/features/'
trainDir = featureDir + 'train/'
trainLabels = np.load(trainDir + 'labels.npy')
classes = sorted(set(trainLabels))
# Get lexicon information
lexicon = create_lexicon.create()
lexicon_means_stds = create_lexicon_means_stds.create()
if globalLexicon == True:
stateProbs = create_probTables.create_stateProbs(lexicon)
transProbs = create_probTables.create_transProbs(lexicon)
# Recognition accuracy names
correct = 0
false = 0
inLex = 0
avgReduction = 0
# Recognize the words
xml = ET.parse(words_file_name).getroot()
recog = Recognizer(sentenceDir, wordDir, xml, img)
for word, word_img in recog.next_word():
required_word = word.get('text')
logging.info("Word:\t\t%s" % required_word)
cuts = recog.find_cuts(word_img)
if cuts is not None:
reduced_lexicon, reduction = reduce_lexicon(cuts, word_img, lexicon, lexicon_means_stds)
avgReduction += reduction
if globalLexicon == False:
stateProbs = create_probTables.create_stateProbs(reduced_lexicon)
transProbs = create_probTables.create_transProbs(reduced_lexicon)
cuts.insert(0, 0) # Also create a window at the start of the word
estimate = recog.recursiveRecognize(
word_img, cuts, reduced_lexicon, stateProbs, transProbs,
classes, minCutWindow, maxCutWindow
)
logging.info("Estimate:\t%s" % estimate)
if required_word in reduced_lexicon:
inLex += 1
if required_word == estimate:
correct += 1
else:
false += 1
accuracy = float(correct)/(correct+false) * 100
logging.info("Correct: %s\tFalse: %s\tAccuracy: %s\n" % (correct, false, accuracy) )
else:
continue
accuracy = float(correct)/(correct+false) * 100
totalInLex = float(inLex)/(correct+false) * 100
avgReduction = avgReduction/float(correct+false)
if state == "external":
logging.info("Complete results for file: %s, Correct: %s\tFalse: %s \tAccuracy: %s" % (sys.argv[1], correct, false, accuracy) )
ET.ElementTree(recog.words).write(sys.argv[3])
else:
return (correct, false)
logging.info("Complete results for file: Correct: %s\tFalse: %s \tAccuracy: %s" % (correct, false, accuracy) )
