def generate_chinese_images_to_check(obj_size=CHAR_IMG_SIZE,
augmentation=False):
print("Get font_file_list ...")
font_file_list = [
os.path.join(FONT_FILE_DIR, font_name)
for font_name in os.listdir(FONT_FILE_DIR)
if font_name.lower()[-4:] in (".otf", ".ttf", ".ttc", ".fon")
]
# font_file_list = [os.path.join(FONT_FINISHED_DIR, "chinese_fonts_暂时移出/康熙字典体完整版本.otf")]
chinese_char_num = len(CHAR2ID_DICT)
total_num = len(font_file_list) * chinese_char_num
count = 0
for font_file in font_file_list: # 外层循环是字体
font_name = os.path.basename(font_file)
font_type = font_name.split(".")[0]
# 创建保存该字体图片的目录
font_img_dir = os.path.join(CHAR_IMGS_DIR, font_type)
remove_then_makedirs(font_img_dir)
for chinese_char, bigger_PIL_img in generate_all_chinese_images_bigger(
font_file, image_size=int(obj_size * 1.2)): # 内层循环是字
# 检查生成的灰度图像是否可用，黑底白字
image_data = list(bigger_PIL_img.getdata())
if sum(image_data) < 10:
continue
if not augmentation:
PIL_img = get_standard_image(bigger_PIL_img,
obj_size,
reverse_color=True)
else:
PIL_img = get_augmented_image(bigger_PIL_img,
obj_size,
rotation=True,
dilate=False,
erode=True,
reverse_color=True)
# 保存生成的字体图片
image_name = chinese_char + ".jpg"
save_path = os.path.join(font_img_dir, image_name)
PIL_img.save(save_path, format="jpeg")
# 当前进度
count += 1
if count % 200 == 0:
print("Progress bar: %.2f%%" % (count * 100 / total_num))
sys.stdout.flush()
return
def convert_chinese_images_to_check(obj_size=CHAR_IMG_SIZE, augmentation=True):
print("Get total images num ...")
font_images_num_list = [
len(os.listdir(os.path.join(EXTERNEL_IMAGES_DIR, content)))
for content in os.listdir(EXTERNEL_IMAGES_DIR)
if os.path.isdir(os.path.join(EXTERNEL_IMAGES_DIR, content))
]
print("Begin to convert images ...")
total_num = sum(font_images_num_list)
count = 0
for font_type, image_paths_list in get_external_image_paths(
root_dir=EXTERNEL_IMAGES_DIR):
# 创建保存该字体图片的目录
font_img_dir = os.path.join(CHAR_IMGS_DIR, font_type)
remove_then_makedirs(font_img_dir)
for image_path in image_paths_list:
# 加载外部图片，将图片调整为正方形
# 为了保证图片旋转时不丢失信息，生成的图片比本来的图片稍微bigger
# 为了方便图片的后续处理，图片必须加载为黑底白字，可以用reverse_color来调整
try:
bigger_PIL_img = load_external_image_bigger(
image_path, white_background=True, reverse_color=True)
except OSError:
print("The image %s result in OSError !" % image_path)
continue
if not augmentation:
PIL_img = get_standard_image(bigger_PIL_img,
obj_size,
reverse_color=True)
else:
PIL_img = get_augmented_image(bigger_PIL_img,
obj_size,
rotation=True,
dilate=False,
erode=True,
reverse_color=True)
# 保存生成的字体图片
image_name = os.path.basename(image_path).split(".")[0] + ".jpg"
save_path = os.path.join(font_img_dir, image_name)
PIL_img.save(save_path, format="jpeg")
# 当前进度
count += 1
if count % 200 == 0:
print("Progress bar: %.2f%%" % (count * 100 / total_num))
sys.stdout.flush()
def generate_chinese_images(obj_size=CHAR_IMG_SIZE,
num_imgs_per_font=NUM_IMAGES_PER_FONT):
print("Get font_file_list ...")
font_file_list = [
os.path.join(FONT_FILE_DIR, font_name)
for font_name in os.listdir(FONT_FILE_DIR)
if font_name.lower()[-4:] in (".otf", ".ttf", ".ttc", ".fon")
]
print("Begin to generate images ...")
chinese_char_num = len(CHAR2ID_DICT)
total_num = len(font_file_list) * chinese_char_num
count = 0
for font_file in font_file_list: # 外层循环是字体
font_name = os.path.basename(font_file)
font_type = font_name.split(".")[0]
# 创建保存该字体图片的目录
save_dir = os.path.join(CHAR_IMGS_DIR, font_type)
remove_then_makedirs(save_dir)
for chinese_char, bigger_PIL_img in generate_all_chinese_images_bigger(
font_file, image_size=int(obj_size * 1.2)): # 内层循环是字
# 检查生成的灰度图像是否可用，黑底白字
image_data = list(bigger_PIL_img.getdata())
if sum(image_data) < 10:
continue
PIL_img_list = \
[get_augmented_image(bigger_PIL_img, obj_size, rotation=True, dilate=False, erode=True, reverse_color=True)
for i in range(num_imgs_per_font)]
# 保存生成的字体图片
for index, PIL_img in enumerate(PIL_img_list):
image_name = chinese_char + "_" + str(index) + ".jpg"
save_path = os.path.join(save_dir, image_name)
PIL_img.save(save_path, format="jpeg")
# 当前进度
count += 1
if count % 200 == 0:
print("Progress bar: %.2f%%" % (count * 100 / total_num))
sys.stdout.flush()
def convert_tfrecords(obj_size=CHAR_IMG_SIZE,
num_imgs_per_font=NUM_IMAGES_PER_FONT):
print("Get total images num ...")
font_images_num_list = [
len(os.listdir(os.path.join(EXTERNEL_IMAGES_DIR, content)))
for content in os.listdir(EXTERNEL_IMAGES_DIR)
if os.path.isdir(os.path.join(EXTERNEL_IMAGES_DIR, content))
]
# 创建保存tfrecords文件的目录
check_or_makedirs(CHAR_TFRECORDS_DIR)
# 可以把变换的图片直接存入tfrecords文件
# 不必将变换的图片先保存到磁盘，再从磁盘读取出来保存到tfrecords文件，这样效率太低
# 通常是用一种字体的一个字图片增强出很多个图片，这些图片最好是分开存放
# 若直接把同一字体同一个字图片增强出的多张图片连续放到同一个tfrecords里，那么每一个训练batch的多样性就不好
writers_list = \
[tf.io.TFRecordWriter(os.path.join(CHAR_TFRECORDS_DIR, "chinese_imgs_%d_from_img.tfrecords" % i))
for i in range(20)]
print("Begin to convert images ...")
total_num = sum(font_images_num_list)
count = 0
for font_type, image_paths_list in get_external_image_paths(
root_dir=EXTERNEL_IMAGES_DIR):
for image_path in image_paths_list:
chinese_char = os.path.basename(image_path)[0]
# 加载外部图片，将图片调整为正方形
# 为了保证图片旋转时不丢失信息，生成的图片比本来的图片稍微bigger
# 为了方便图片的后续处理，图片必须加载为黑底白字，可以用reverse_color来调整
try:
bigger_PIL_img = load_external_image_bigger(
image_path, white_background=True, reverse_color=True)
except OSError:
print("The image %s result in OSError !" % image_path)
continue
PIL_img_list = \
[get_augmented_image(bigger_PIL_img, obj_size, rotation=True, dilate=False, erode=True, reverse_color=True)
for i in range(num_imgs_per_font)]
# 保存生成的字体图片
for index, PIL_img in enumerate(PIL_img_list):
# train_set和test_set的比例约为 5:1
writer = random.choice(writers_list)
bytes_image = PIL_img.tobytes() # 将图片转化为原生bytes
bytes_char = chinese_char.encode('utf-8')
example = tf.train.Example(features=tf.train.Features(
feature={
'bytes_image':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[bytes_image])),
'bytes_char':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[bytes_char])),
}))
writer.write(example.SerializeToString())
# 当前进度
count += 1
if count % 200 == 0:
print("Progress bar: %.2f%%" % (count * 100 / total_num))
sys.stdout.flush()
# 关闭所有的 tfrecords writer
[writer.close() for writer in writers_list]
def generate_tfrecords(obj_size=CHAR_IMG_SIZE,
num_imgs_per_font=NUM_IMAGES_PER_FONT):
print("Get font_file_list ...")
font_file_list = [
os.path.join(FONT_FILE_DIR, font_name)
for font_name in os.listdir(FONT_FILE_DIR)
if font_name.lower()[-4:] in (".otf", ".ttf", ".ttc", ".fon")
]
# 创建保存tfrecords文件的目录
check_or_makedirs(CHAR_TFRECORDS_DIR)
# 可以把生成的图片直接存入tfrecords文件
# 不必将生成的图片先保存到磁盘，再从磁盘读取出来保存到tfrecords文件，这样效率太低
# 通常是用某种字体对一个字生成很多个增强的图片，这些图片最好是分开存放
# 若直接把同一字体同一个字的多张图片连续放到同一个tfrecords里，那么训练batch的多样性不好
writers_list = \
[tf.io.TFRecordWriter(os.path.join(CHAR_TFRECORDS_DIR, "chinese_imgs_%d_from_font.tfrecords" % i))
for i in range(20)]
print("Begin to generate images ...")
chinese_char_num = len(CHAR2ID_DICT)
total_num = len(font_file_list) * chinese_char_num
count = 0
for font_file in font_file_list: # 外层循环是字体
for chinese_char, bigger_PIL_img in generate_all_chinese_images_bigger(
font_file, image_size=int(obj_size * 1.2)): # 内层循环是字
# 检查生成的灰度图像是否可用，黑底白字
image_data = list(bigger_PIL_img.getdata())
if sum(image_data) < 10:
continue
PIL_img_list = \
[get_augmented_image(bigger_PIL_img, obj_size, rotation=True, dilate=False, erode=True, reverse_color=True)
for i in range(num_imgs_per_font)]
# 保存生成的字体图片
for PIL_img in PIL_img_list:
writer = random.choice(writers_list)
bytes_image = PIL_img.tobytes() # 将图片转化为原生bytes
bytes_char = chinese_char.encode('utf-8')
example = tf.train.Example(features=tf.train.Features(
feature={
'bytes_image':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[bytes_image])),
'img_height':
tf.train.Feature(int64_list=tf.train.Int64List(
value=[PIL_img.height])),
'img_width':
tf.train.Feature(int64_list=tf.train.Int64List(
value=[PIL_img.width])),
'bytes_char':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[bytes_char]))
}))
writer.write(example.SerializeToString())
# 当前进度
count += 1
if count % 200 == 0:
print("Progress bar: %.2f%%" % (count * 100 / total_num))
sys.stdout.flush()
# 关闭所有的tfrecords写者
[writer.close() for writer in writers_list]
