def generate(cls, index, text, font, out_dir, size, extension,
skewing_angle, random_skew, blur, random_blur,
background_type, distorsion_type, distorsion_orientation,
is_handwritten, name_format, width, alignment, text_color,
orientation, space_width, margins, fit):
image = None
text = get_display(text)
margin_top, margin_left, margin_bottom, margin_right = margins
horizontal_margin = margin_left + margin_right
vertical_margin = margin_top + margin_bottom
##########################
# Create picture of text #
##########################
if is_handwritten:
if orientation == 1:
raise ValueError("Vertical handwritten text is unavailable")
image = handwritten_text_generator.generate(text, text_color, fit)
else:
image = computer_text_generator.generate(text, font, text_color,
size, orientation,
space_width, fit)
random_angle = random.randint(0 - skewing_angle, skewing_angle)
rotated_img = image.rotate(
skewing_angle if not random_skew else random_angle, expand=1)
#############################
# Apply distorsion to image #
#############################
if distorsion_type == 0:
distorted_img = rotated_img # Mind = blown
elif distorsion_type == 1:
distorted_img = distorsion_generator.sin(
rotated_img,
vertical=(distorsion_orientation == 0
or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1
or distorsion_orientation == 2))
elif distorsion_type == 2:
distorted_img = distorsion_generator.cos(
rotated_img,
vertical=(distorsion_orientation == 0
or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1
or distorsion_orientation == 2))
else:
distorted_img = distorsion_generator.random(
rotated_img,
vertical=(distorsion_orientation == 0
or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1
or distorsion_orientation == 2))
##################################
# Resize image to desired format #
##################################
# Horizontal text
if orientation == 0:
new_width = int(
distorted_img.size[0] *
(float(size - vertical_margin) / float(distorted_img.size[1])))
resized_img = distorted_img.resize(
(new_width, size - vertical_margin), Image.ANTIALIAS)
background_width = width if width > 0 else new_width + horizontal_margin
background_height = size
# Vertical text
elif orientation == 1:
new_height = int(
float(distorted_img.size[1]) *
(float(size - horizontal_margin) /
float(distorted_img.size[0])))
resized_img = distorted_img.resize(
(size - horizontal_margin, new_height), Image.ANTIALIAS)
background_width = size
background_height = new_height + vertical_margin
else:
raise ValueError("Invalid orientation")
#############################
# Generate background image #
#############################
if background_type == 0:
background = background_generator.gaussian_noise(
background_height, background_width)
elif background_type == 1:
background = background_generator.plain_white(
background_height, background_width)
elif background_type == 2:
background = background_generator.quasicrystal(
background_height, background_width)
else:
background = background_generator.picture(background_height,
background_width)
#############################
# Place text with alignment #
#############################
new_text_width, _ = resized_img.size
if alignment == 0 or width == -1:
background.paste(resized_img, (margin_left, margin_top),
resized_img)
elif alignment == 1:
background.paste(
resized_img,
(int(background_width / 2 - new_text_width / 2), margin_top),
resized_img)
else:
background.paste(
resized_img,
(background_width - new_text_width - margin_right, margin_top),
resized_img)
##################################
# Apply gaussian blur #
##################################
final_image = background.filter(
ImageFilter.GaussianBlur(
radius=(blur if not random_blur else random.randint(0, blur))))
#####################################
# Generate name for resulting image #
#####################################
text = get_display(text)
if name_format == 0:
image_name = '{}.{}'.format(text, extension)
elif name_format == 1:
image_name = '{}.{}'.format(text, extension)
elif name_format == 2:
image_name = '{}.{}'.format(text, extension)
else:
print('{} is not a valid name format. Using default.'.format(
name_format))
image_name = '{}_{}.{}'.format(text, str(index), extension)
# Save the image
final_image.convert('RGB').save(os.path.join(out_dir, image_name))
def generate(cls, index, text, font, out_dir, size, extension, skewing_angle, random_skew, blur, random_blur, background_type, distorsion_type, distorsion_orientation, is_handwritten, name_format, width, alignment, text_color, orientation, space_width, margins, fit):
num_zero = 6 - len(str(index))
image_name = '{}.{}'.format('0'*num_zero+str(index),extension)
file = open("bb_result.txt","a+")
file.write("\n"+image_name.split(".")[0] + " " +text.replace(" ", ""))
file.close
image = None
size_percentage = rnd.randrange(75,125,5)
size = int(size * (size_percentage/100))
margin_top, margin_left, margin_bottom, margin_right = margins
horizontal_margin = margin_left + margin_right
vertical_margin = margin_top + margin_bottom
# print("margins is :" , margins)
##########################
# Create picture of text #
##########################
if is_handwritten:
if orientation == 1:
raise ValueError("Vertical handwritten text is unavailable")
image = handwritten_text_generator.generate(text, text_color, fit)
else:
image = computer_text_generator.generate(text, font, text_color, size, orientation, space_width, fit)
random_angle = rnd.randint(0-skewing_angle, skewing_angle)
rotated_img = image.rotate(skewing_angle if not random_skew else random_angle, expand=1)
#############################
# Apply distorsion to image #
#############################
if distorsion_type == 0:
distorted_img = rotated_img # Mind = blown
elif distorsion_type == 1:
distorted_img = distorsion_generator.sin(
rotated_img,
vertical=(distorsion_orientation == 0 or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1 or distorsion_orientation == 2)
)
elif distorsion_type == 2:
distorted_img = distorsion_generator.cos(
rotated_img,
vertical=(distorsion_orientation == 0 or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1 or distorsion_orientation == 2)
)
else:
distorted_img = distorsion_generator.random(
rotated_img,
vertical=(distorsion_orientation == 0 or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1 or distorsion_orientation == 2)
)
##################################
# Resize image to desired format #
##################################
# Horizontal text
if orientation == 0:
new_width = int(distorted_img.size[0] * (float(size - vertical_margin) / float(distorted_img.size[1])))
resized_img = distorted_img.resize((new_width, size - vertical_margin), Image.ANTIALIAS)
background_width = width if width > 0 else new_width + horizontal_margin
background_height = size
# Vertical text
elif orientation == 1:
new_height = int(float(distorted_img.size[1]) * (float(size - horizontal_margin) / float(distorted_img.size[0])))
resized_img = distorted_img.resize((size - horizontal_margin, new_height), Image.ANTIALIAS)
background_width = size
background_height = new_height + vertical_margin
else:
raise ValueError("Invalid orientation")
#############################
# Generate background image #
#############################
if background_type == 0:
background = background_generator.gaussian_noise(background_height, background_width)
elif background_type == 1:
background = background_generator.plain_white(background_height, background_width)
elif background_type == 2:
background = background_generator.quasicrystal(background_height, background_width)
else:
background = background_generator.picture(background_height, background_width)
#############################
# Place text with alignment #
#############################
new_text_width, _ = resized_img.size
# print("resized_img.size is : ", resized_img.size)
if alignment == 0 or width == -1:
background.paste(resized_img, (margin_left, margin_top), resized_img)
elif alignment == 1:
background.paste(resized_img, (int(background_width / 2 - new_text_width / 2), margin_top), resized_img)
x0 = int(background_width / 2 - new_text_width / 2 )
y0 = margin_top
x1 = x0 + resized_img.size[0]
y1 = y0 + resized_img.size[1]
number =int((len(text)+1)/2)
# print("digits number in text is : ", number)
# print("text is : ", text)
angle = math.radians(random_angle)
# print("random_angle is ", random_angle)
#digit_width = int(((x1-x0)/number)*(1+0.005*(number)))
digit_width = int((1+angle/90)*(new_text_width/number))
for i in range(number):
x00 = x0 + i * digit_width+ 0.7*(1.5*i-number/2)
x11 = x0 + (i+1) * (digit_width)+ 0.7*(1.5*i-number/2)
txt_draw = ImageDraw.Draw(background)
x_center = 0.5 * background.size[0]
y_center = 0.5 * background.size[1]
# test
rx0 = x00
ry0 = (y0-y_center)*math.cos(angle) - (x00-x_center)*math.sin(angle) + y_center
rx1 = x11
ry1 = (y1-y_center)*math.cos(angle) - (x11-x_center)*math.sin(angle) + y_center
# rx0 = (x00-x_center)*math.cos(angle) + (y0-y_center)*math.sin(angle) + x_center
# ry0 = (y0-y_center)*math.cos(angle) - (x00-x_center)*math.sin(angle) + y_center
#
# rx1 = (x11-x_center)*math.cos(angle) + (y1-y_center)*math.sin(angle) + x_center
# ry1 = (y1-y_center)*math.cos(angle) - (x11-x_center)*math.sin(angle) + y_center
# the rotated bounding box for each digit is (rx0,ry0,rx1,ry1)
################################# txt_draw.rectangle((rx0,ry0,rx1,ry1), fill=None, outline=None, width=1)
# print("the rotated bounding box for each digit is ", rx0,ry0,rx1,ry1)
file = open("bb_result.txt","a+")
# file.write("\nbbox coordinates : x_min: "+ str(rx0) +" y_min: "+str(ry0)+" x_max: "+ str(rx1)+" y_max: "+str(ry1))
file.write(" "+ str(rx0)+" " + str(ry0)+" " +str(rx1)+" "+str(ry1))
file.close
else:
background.paste(resized_img, (background_width - new_text_width - margin_right, margin_top), resized_img)
##################################
# Apply gaussian blur #
##################################
final_image = background.filter(
ImageFilter.GaussianBlur(
radius=(blur if not random_blur else rnd.randint(0, blur))
)
)
# Save the image
final_image.convert('RGB').save(os.path.join(out_dir, image_name))
final_img_copy =np.array(final_image.convert('RGB'))
file = open("bb_result.txt","a+")
file.write(" "+str(final_img_copy.shape[1])+" "+str(final_img_copy.shape[0]))
file.close
def generate(cls, ithread, index, text, word_labels, font, out_dir,
save4cycleGAN, tf_flag, tfr_out_dir, writer, size, extension,
skewing_angle, random_skew, blur, random_blur,
background_type, distorsion_type, perspective_type,
distorsion_orientation, is_handwritten, name_format, width,
alignment, text_color, orientation, space_width, margins, fit,
trte):
image = None
margin_top, margin_left, margin_bottom, margin_right = margins
horizontal_margin = margin_left + margin_right
vertical_margin = margin_top + margin_bottom
##########################
# Create picture of text #
##########################
if is_handwritten:
if orientation == 1:
raise ValueError("Vertical handwritten text is unavailable")
image = handwritten_text_generator.generate(text, text_color, fit)
else:
image = computer_text_generator.generate(text, font, text_color,
size, orientation,
space_width, fit)
random_angle = random.randint(0 - skewing_angle, skewing_angle)
rotated_img = image.rotate(
skewing_angle if not random_skew else random_angle, expand=1)
# plt.imshow(rotated_img)
# plt.show()
##################################
# Apply perspective to image #
##################################
if perspective_type == 0:
perspectived_img = rotated_img
elif perspective_type == 1:
perspectived_img = perspective_generator._apply_func_perspective(
rotated_img)
perspectived_img = perspectived_img
#############################
# Apply distorsion to image #
#############################
if distorsion_type == 0:
distorted_img = perspectived_img # Mind = blown
elif distorsion_type == 1:
distorted_img = distorsion_generator.sin(
perspectived_img,
vertical=(distorsion_orientation == 0
or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1
or distorsion_orientation == 2))
elif distorsion_type == 2:
distorted_img = distorsion_generator.cos(
perspectived_img,
vertical=(distorsion_orientation == 0
or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1
or distorsion_orientation == 2))
else:
distorted_img = distorsion_generator.random(
perspectived_img,
vertical=(distorsion_orientation == 0
or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1
or distorsion_orientation == 2))
# plt.imshow(distorted_img)
# plt.show()
##################################
# Resize image to desired format #
##################################
# Horizontal text
if orientation == 0:
new_width = int(
distorted_img.size[0] *
(float(size - vertical_margin) / float(distorted_img.size[1])))
resized_img = distorted_img.resize(
(new_width, size - vertical_margin), Image.ANTIALIAS)
background_width = width if width > 0 else new_width + horizontal_margin
background_height = size
# Vertical text
elif orientation == 1:
new_height = int(
float(distorted_img.size[1]) *
(float(size - horizontal_margin) /
float(distorted_img.size[0])))
resized_img = distorted_img.resize(
(size - horizontal_margin, new_height), Image.ANTIALIAS)
background_width = size
background_height = new_height + vertical_margin
else:
raise ValueError("Invalid orientation")
#############################
# Generate background image #
#############################
if background_type == 0:
background = background_generator.gaussian_noise(
background_height, background_width)
elif background_type == 1:
background = background_generator.plain_white(
background_height, background_width)
elif background_type == 2:
background = background_generator.quasicrystal(
background_height, background_width)
elif background_type == 3:
background = background_generator.picture(background_height,
background_width)
#############################
# Place text with alignment #
#############################
new_text_width, _ = resized_img.size
if alignment == 0 or width == -1:
background.paste(resized_img, (margin_left, margin_top),
resized_img)
elif alignment == 1:
background.paste(
resized_img,
(int(background_width / 2 - new_text_width / 2), margin_top),
resized_img)
else:
background.paste(
resized_img,
(background_width - new_text_width - margin_right, margin_top),
resized_img)
##################################
# Apply gaussian blur #
##################################
final_image = background.filter(
ImageFilter.GaussianBlur(
radius=(blur if not random_blur else random.randint(0, blur))))
#####################################
# Generate name for resulting image #
#####################################
if name_format == 0:
image_name = '{}_{}_{}.{}'.format(text, str(index),
str(font.split('/')[-1]),
extension)
elif name_format == 1:
image_name = '{}_{}.{}'.format(str(index), text, extension)
elif name_format == 2:
image_name = '{}.{}'.format(str(index), extension)
elif name_format == 3:
image_name = '{}_{}_{}.{}'.format(text, str(index),
str(font.split('/')[-1]),
extension)
else:
print('{} is not a valid name format. Using default.'.format(
name_format))
image_name = '{}_{}.{}'.format(text, str(index), extension)
final_image
# tf record writer
# init for the first time
# Save the image
# final_image.convert('RGB').save(os.path.join(out_dir, image_name))
# img_OpenCV_rawBGR = cv2.cvtColor(np.asarray(final_image.convert('RGB')), cv2.COLOR_RGB2BGR)
# img_OpenCV_rawRGB = np.asarray(final_image.convert('RGB'))
img_OpenCV_rawGRAY = cv2.cvtColor(
np.asarray(final_image.convert('RGB')), cv2.COLOR_RGB2GRAY)
img_OpenCV = img_OpenCV_rawGRAY
plt.imshow(img_OpenCV)
plt.show()
# cv2.imshow("img_OpenCV_rawBGR",img_OpenCV_rawBGR)
# cv2.imshow("img_OpenCV_rawRGB",img_OpenCV_rawRGB)
# cv2.imshow("img_OpenCV_rawGRAY",img_OpenCV_rawGRAY)
# cv2.waitKey(0)
# img_OpenCV = rgb2gray(img_OpenCV_raw)
if orientation == 1:
im_size = img_OpenCV.shape
img_re_width = im_size[1]
# img_re_width = int(im_size[1] * img_re_height * 1.0 / im_size[0])
# img_OpenCV = cv2.resize(img_OpenCV, (img_re_height, img_re_width))
else:
im_size = img_OpenCV.shape
img_re_width = int(im_size[1] * img_re_height * 1.0 / im_size[0])
img_OpenCV = cv2.resize(img_OpenCV, (img_re_width, img_re_height))
if name_format == 3:
if not os.path.exists(out_dir + "/{}_{}".format(trte, ithread)):
os.makedirs(out_dir + "/{}_{}".format(trte, ithread))
name = out_dir + "/{}_{}/tmp-{}—{}-font{}.jpg".format(
trte, ithread, index, word_labels, str(font.split('/')[-1]))
if save4cycleGAN:
img_OpenCV = cv2.resize(img_OpenCV, (256, 256))
cv2.imwrite(name, img_OpenCV)
if tf_flag == 1:
img = tf.gfile.GFile(name, 'rb').read()
write_one_data(writer, img, word_labels, img_re_width)
def generate(cls, index, text, font, out_dir, size, extension, channel,
skewing_angle, random_skew, blur, random_blur,
background_type, distorsion_type, distorsion_orientation,
random_process, noise, erode, dilate, incline,
is_handwritten, name_format, width, alignment, text_color,
orientation, space_width, margins, fit):
image = None
margin_top, margin_left, margin_bottom, margin_right = margins
horizontal_margin = margin_left + margin_right
vertical_margin = margin_top + margin_bottom
##########################
# Create picture of text #
##########################
if is_handwritten:
if orientation == 1:
raise ValueError("Vertical handwritten text is unavailable")
image = handwritten_text_generator.generate(text, text_color, fit)
else:
image = computer_text_generator.generate(text, font, text_color, size, orientation, space_width, fit)
random_angle = random.random()*random.randint(0-skewing_angle, skewing_angle)
skewing_angle = skewing_angle if not random_skew else random_angle
rotated_img = image.rotate(skewing_angle, expand=1)
#############################
# Apply distorsion to image #
#############################
if distorsion_type == 0:
distorted_img = rotated_img # Mind = blown
elif distorsion_type == 1:
distorted_img = distorsion_generator.sin(
rotated_img,
vertical=(distorsion_orientation == 0 or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1 or distorsion_orientation == 2)
)
elif distorsion_type == 2:
distorted_img = distorsion_generator.cos(
rotated_img,
vertical=(distorsion_orientation == 0 or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1 or distorsion_orientation == 2)
)
else:
distorted_img = distorsion_generator.random(
rotated_img,
vertical=(distorsion_orientation == 0 or distorsion_orientation == 2),
horizontal=(distorsion_orientation == 1 or distorsion_orientation == 2)
)
##################################
# Resize image to desired format #
##################################
# Horizontal text
if orientation == 0:
new_width = int(distorted_img.size[0] * (float(size - vertical_margin) / float(distorted_img.size[1])))
resized_img = distorted_img.resize((new_width, size - vertical_margin), Image.ANTIALIAS)
background_width = width if width > 0 else new_width + horizontal_margin
background_height = size
# Vertical text
elif orientation == 1:
new_height = int(float(distorted_img.size[1]) * (float(size - horizontal_margin) / float(distorted_img.size[0])))
resized_img = distorted_img.resize((size - horizontal_margin, new_height), Image.ANTIALIAS)
background_width = size
background_height = new_height + vertical_margin
else:
raise ValueError("Invalid orientation")
#############################
# Generate background image #
#############################
if background_type == 0:
background = background_generator.gaussian_noise(background_height, background_width)
elif background_type == 1:
background = background_generator.plain_white(background_height, background_width)
elif background_type == 2:
background = background_generator.quasicrystal(background_height, background_width)
else:
background = background_generator.picture(background_height, background_width)
#############################
# Place text with alignment #
#############################
new_text_width, _ = resized_img.size
if alignment == 0 or width == -1:
background.paste(resized_img, (margin_left, margin_top), resized_img)
elif alignment == 1:
background.paste(resized_img, (int(background_width / 2 - new_text_width / 2), margin_top), resized_img)
else:
background.paste(resized_img, (background_width - new_text_width - margin_right, margin_top), resized_img)
##################################
# Apply gaussian blur #
##################################
blur = blur if not random_blur else random.random()*blur
final_image = background.filter(
ImageFilter.GaussianBlur(
radius = blur
)
)
##################################
# Apply image process #
##################################
final_image = np.asarray(final_image)
incline = min(incline, int(final_image.shape[1]/4))
final_image, noise, dilate, erode, incline = image_process.do(random_process, noise, erode, dilate, incline, final_image)
final_image = Image.fromarray(final_image)
#####################################
# Generate name for resulting image #
#####################################
if name_format == 0:
image_name = '{}_{}.{}'.format(text, str(index), extension)
elif name_format == 1:
image_name = '{}_{}.{}'.format(str(index), text, extension)
elif name_format == 2:
image_name = '{}.{}'.format(str(index),extension)
else:
print('{} is not a valid name format. Using default.'.format(name_format))
image_name = '{}_{}.{}'.format(text, str(index), extension)
# Save the image
if channel == 3:
final_image = final_image.convert('RGB')
elif channel ==1:
final_image = final_image.convert('L')
#final_image.save(os.path.join(out_dir, image_name))
# Resize
final_image = np.asarray(final_image)
if orientation == 0:
final_image = cv2.resize(final_image, (int(final_image.shape[1] * 32 / final_image.shape[0]), 32))
else:
final_image = cv2.resize(final_image, (32, int(final_image.shape[0] * 32 / final_image.shape[1])))
if platform.system() == "Windows":
font = font.split('\\')[1].split('.')[0]
else: # linux
font = font.split('/')[2].split('.')[0]
background = ['Guassian', 'Plain white', 'Quasicrystal','Image']
distorsion = ['None', 'Sine wave', 'Cosine wave']
data = pd.DataFrame(np.array([[index, text, len(text), final_image.shape[1], size, font, skewing_angle, blur,
distorsion[distorsion_type], background[background_type],
noise, erode, dilate, incline]]),
columns=['index', 'text', 'text_length', 'img_shape', 'font_size', 'font_id', 'skew_angle',
'blur', 'distorsion_type', 'background_type','num_noise','erode','dilate','incline'])
if channel == 1 :
final_image = np.asarray(torch.Tensor(np.asarray(final_image)).unsqueeze(0).permute(1,2,0),dtype='uint8')
return data, final_image