def test(self, image, char=None):
coms = connected_components(image)
com = largest_component(coms)
x = params_from_component(com, with_one=False)
indexes, errors = self.kdtree.knn(x, self.k)
scores = {lab: 0.0 for lab in self.label_set}
for i in indexes[0]:
scores[self.labels[i]] += 1.0 / self.k
del scores['']
return KNNClassifications(scores)
def __init__(self, img):
self.img = img
self.binary_img = binary(img, invert=True)
knn_test = KNNTest(data_path)
self.char_test = Combiner([cont_test, mask_test, knn_test.test]).test
self.objs = [
ImgObj(x, self) for x in connected_components(self.binary_img) if self.good_component(x)
]
self.compute_char_scores()
self.group_words(score_threshold=0.2)
self.spell_corrector = SpellingCorrector('/usr/share/dict/words')
def test(image, char=None):
classifications = MLClassifications()
coms = connected_components(image)
com = largest_component(coms)
xs = params_from_component(com, with_one=True)
for c, ws in models.items():
s = dot(ws, Trainer.get_transformed_data(xs, polynomial_transform_order))
print c, s
classifications.add(c, s)
return classifications
def generate_noise_points(self, output, num_samples):
num_noise_points = 0
for noise_img in self.noise_images:
mat = cv.LoadImageM(noise_img, cv.CV_LOAD_IMAGE_GRAYSCALE)
mat = binary(mat, invert=True)
for com in connected_components(mat):
if 10 < com.mask.rows and 5 < com.mask.cols:
num_noise_points += 1
output.write("\t%s\n" % float_list_to_tsv(params_from_component(com)))
if num_samples <= num_noise_points:
return num_samples
return num_noise_points
def generate_data_points(self, output, num_samples):
image_size = 64
font_size = 48
image = Image.new('L', (image_size, image_size), "#000000")
mat = cv.CreateMatHeader(image_size, image_size, cv.CV_8U)
rot = cv.CreateMat(2, 3, cv.CV_32F)
mapmat = cv.CreateMat(image_size, image_size, cv.CV_8U)
points_per_font = num_samples / len(self.fonts)
for f in self.fonts:
font = load_font(f, font_size)
for c in self.chars:
for _ in range(points_per_font):
draw_text(font, c, image)
cv.SetData(mat, image.tostring())
cx = mat.cols / 2.0 + random.gauss(0, mat.cols / 30.0)
cy = mat.rows / 2.0 + random.gauss(0, mat.rows / 30.0)
angle = random.gauss(0, 1 / 15.0)
cv.GetRotationMatrix2D((cx, cy), angle, 1.0, rot)
cv.WarpAffine(mat, mapmat, rot)
coms = connected_components(mapmat)
com = largest_component(coms)
output.write(c + "\t%s\n" % float_list_to_tsv(params_from_component(com)))