def initialize_train_test_folder(file_path):
"""
create new train, test and validation folder if
it doesn't exist.
:param file_path: main_path
"""
train_folder = '{}/{}'.format(file_path, 'train')
test_folder = '{}/{}'.format(file_path, 'test')
valid_folder = '{}/{}'.format(file_path, 'validation')
check_directory(train_folder)
check_directory(test_folder)
check_directory(valid_folder)
def batch_image_process(ori_set, target_set, process, stage=None, num_op=None):
"""
:param ori_set:
:param target_set:
:param process: process can be 'crop' or 'flip'
:return:
"""
check_directory([TARGET_DATA_DIR, TEST_TARGET_DATA_DIR, TRAIN_TARGET_DATA_DIR, VALID_TARGET_DATA_DIR])
check_directory([CROP_TARGET_DATA_DIR, CROP_TEST_TARGET_DATA_DIR, CROP_TRAIN_TARGET_DATA_DIR, CROP_VALID_TARGET_DATA_DIR])
for file in os.listdir(ori_set):
print(file)
image_file = os.path.join(ori_set, str(file), str(file) + '.bmp')
det_mask_file = os.path.join(ori_set, str(file), str(file) + '_detection.bmp')
cls_mask_file = os.path.join(ori_set, str(file), str(file) + '_classification.bmp')
image = cv2.imread(image_file)
det_mask = cv2.imread(det_mask_file)
cls_mask = cv2.imread(cls_mask_file)
if process == 'crop':
image = cv2.resize(image, (512, 512))
det_mask = cv2.resize(det_mask, (512, 512))
cls_mask = cv2.resize(cls_mask, (512, 512))
crop_list = crop_image_parts(image, det_mask, cls_mask)
list_file_create = [os.path.join(target_set, str(file)+'_1'),
os.path.join(target_set, str(file)+'_2'),
os.path.join(target_set, str(file)+'_3'),
os.path.join(target_set, str(file)+'_4')]
check_directory(list_file_create)
list_img_create = [os.path.join(target_set, str(file)+ '_1', str(file)+'_1.bmp'),
os.path.join(target_set, str(file)+ '_2', str(file)+'_2.bmp'),
os.path.join(target_set, str(file)+ '_3', str(file)+'_3.bmp'),
os.path.join(target_set, str(file)+'_4', str(file)+'_4.bmp'),
os.path.join(target_set, str(file)+'_1',str(file)+'_1_detection.bmp'),
os.path.join(target_set, str(file)+'_2',str(file)+'_2_detection.bmp'),
os.path.join(target_set, str(file)+'_3',str(file)+'_3_detection.bmp'),
os.path.join(target_set, str(file)+'_4',str(file)+'_4_detection.bmp'),
os.path.join(target_set, str(file) + '_1', str(file) + '_1_classification.bmp'),
os.path.join(target_set, str(file) + '_2', str(file) + '_2_classification.bmp'),
os.path.join(target_set, str(file) + '_3', str(file) + '_3_classification.bmp'),
os.path.join(target_set, str(file) + '_4', str(file) + '_4_classification.bmp'),
os.path.join(target_set, str(file) + '_1', str(file) + '_1_original.bmp'),
os.path.join(target_set, str(file) + '_2', str(file) + '_2_original.bmp'),
os.path.join(target_set, str(file) + '_3', str(file) + '_3_original.bmp'),
os.path.join(target_set, str(file) + '_4', str(file) + '_4_original.bmp')]
for order, img in enumerate(crop_list):
check_cv2_imwrite(list_img_create[order], img)
elif process == 'flip':
aug_lrimage, aug_lrdetmask, aug_lrclsmask = img_aug(image, det_mask, cls_mask, 'fliplr')
aug_upimage, aug_updetmask, aug_upclsmask = img_aug(image, det_mask, cls_mask, 'flipup')
list_file_create = [os.path.join(target_set, str(file) + '_lr'),
os.path.join(target_set, str(file) + '_up')]
check_directory(list_file_create)
list_img_create = [os.path.join(target_set, str(file) + '_lr', str(file) + '_lr.bmp'),
os.path.join(target_set, str(file) + '_up', str(file) + '_up.bmp'),
os.path.join(target_set, str(file) + '_lr', str(file) + '_lr_detection.bmp'),
os.path.join(target_set, str(file) + '_up', str(file) + '_up_detection.bmp'),
os.path.join(target_set, str(file) + '_lr', str(file) + '_lr_classification.bmp'),
os.path.join(target_set, str(file) + '_up', str(file) + '_up_classification.bmp'),
os.path.join(target_set, str(file) + '_lr', str(file) + '_lr_original.bmp'),
os.path.join(target_set, str(file) + '_up', str(file) + '_up_original.bmp')]
flip_list = [aug_lrimage, aug_upimage,
aug_lrdetmask, aug_updetmask,
aug_lrclsmask, aug_upclsmask,
aug_lrimage, aug_upimage]
for order, img in enumerate(flip_list):
check_cv2_imwrite(list_img_create[order], img)
elif process == 'shear' or process == 'channel' or process == 'random' or process=='zoom' or process=='rotate':
aug_image, aug_detmask, aug_clsmask = img_aug(image, det_mask, cls_mask, process)
list_file_create = [os.path.join(target_set, str(file) + '_' + process + str(stage))]
check_directory(list_file_create)
list_img_create = [os.path.join(target_set, str(file) + '_' + process + str(stage), str(file) + '_' + process + str(stage) + '.bmp'),
os.path.join(target_set, str(file) + '_' + process + str(stage), str(file) + '_' + process + str(stage) + '_detection.bmp'),
os.path.join(target_set, str(file) + '_' + process + str(stage), str(file) + '_' + process + str(stage) + '_classification.bmp'),
os.path.join(target_set, str(file) + '_' + process + str(stage), str(file) + '_' + process + str(stage) + '_original.bmp')]
image_list = [aug_image, aug_detmask, aug_clsmask, aug_image]
for order, img in enumerate(image_list):
check_cv2_imwrite(list_img_create[order], img)
else:
raise ValueError('augmentation type wrong, check documents.')
def __init__(self, data_path = None, old_filename = None, new_filename = None):
self.data_path = data_path
self.old_filename = '{}/{}'.format(data_path, old_filename)
self.new_filename = '{}/{}'.format(data_path, new_filename)
check_directory(self.new_filename)
initialize_train_test_folder(self.new_filename)
def persistence_image_data_to_tfrecords(x_data,
y_data,
data_type,
split_index=128):
"""
persistence_image_data_to_tfrecords(x_data, y_data, data_type, split_index)
- Make images into tfrecords and save
- Crop / Resize / Convert rgb into grayscale
"""
TARGET_DIR = _get_target_dir()
OUTPUT_DIR = os.path.join(FLAGS.output_dir, TARGET_DIR, data_type)
IMAGE_DIR = FLAGS.image_dir
myutil.check_directory("./resized_img") # Check directory of "resized_img"
myutil.check_directory("./tfrecords")
myutil.check_directory("./tfrecords/cropping")
myutil.check_directory("./tfrecords/cropping/train")
myutil.check_directory("./tfrecords/cropping/test")
myutil.check_directory(OUTPUT_DIR) # Check directory of OUTPUT_DIR
writer = None
sess = None
current_index = 0
from sklearn import preprocessing
le = preprocessing.LabelEncoder()
le.fit(y_data)
y_data_size = len(le.classes_)
# https://stackoverflow.com/questions/45427637/is-there-a-more-simple-way-to-handle-batch-inputs-from-tfrecords
for images_filename, y_label in zip(x_data, y_data):
if not (images_filename[-3:] == "jpg"):
print("Error - ", images_filename)
continue
if current_index % split_index == 0:
if writer:
writer.close()
if sess:
sess.close()
tf.reset_default_graph()
graph = tf.get_default_graph()
sess = tf.Session(graph=graph)
sess.run(tf.global_variables_initializer())
record_filename = "{output_dir}/{data_type}-{current_index}.tfrecords".format(
output_dir=OUTPUT_DIR,
data_type=data_type,
current_index=current_index)
print("=============>", record_filename)
print("current index : {0}".format(current_index, ))
writer = tf.python_io.TFRecordWriter(record_filename)
file_full_path = os.path.join(IMAGE_DIR, y_label, images_filename)
try:
image_file = tf.read_file(file_full_path)
decoded_img = tf.image.decode_jpeg(
image_file, channels=3) # 'jpg' file decoding
except tf.errors.InvalidArgumentError as e:
print(e)
print("Error : ", images_filename)
continue
image_list = [decoded_img]
# If I wanna crop images
if FLAGS.cropping:
image_list = []
# Get size information
size_info = iputil.get_orginal_size_info(y_label, images_filename)
target_img = tf.image.resize_images(
decoded_img, [size_info["height"], size_info["width"]])
# Get bounding box to crop image (it is list)
# bounding_box = [[xmin,ymin,bounding_width,bounding_height]]
bounding_box = iputil.get_bounding_size_info(
y_label, images_filename)
# Crop image with bounding box
for box in bounding_box:
# tf.img.crop_to_bounding_box
# tf.img.* 함수들은 width보다 height를 먼저 쓴다
# -> 이미지 좌표상 x,y,width,height의 순서를 잘 확인하지 않으면 전처리가 잘못 적용되니 유의할 것!
cropped_img = tf.image.crop_to_bounding_box(
target_img, box[1], box[0], box[3], box[2])
# cropped_img = tf.image.crop_and_resize(decoded_img, bounding_box, i, crop_size=[FLAGS.img_height, FLAGS.img_width])
"""
tf.image.crop_and_resize를 사용하고 싶은데 뭔가 자꾸 오류가 남...
tf.image.crop_and_resize 함수 내에 세팅하는 값의 문제 - 함수 사용법을 좀더 살펴봐야 함!
tf.image.crop_and_resize(
image,
boxes,
box_ind,
crop_size,
method='bilinear',
extrapolation_value=0,
name=None
)
"""
image_list.append(cropped_img)
for image in image_list:
try:
imsave(
"./resized_img/" + images_filename + "_" +
str(current_index) + ".jpeg", sess.run(image))
except OSError as e:
print(e)
gray_img = tf.image.rgb_to_grayscale(image) # rgb to grayscale
resized_img = tf.image.resize_images(
gray_img, [FLAGS.img_width, FLAGS.img_height
]) # Resize image with new width and height
image_bytes = sess.run(tf.cast(resized_img, tf.uint8)).tobytes()
y_data_label = le.transform([y_label])
lbl_one_hot = tf.one_hot(y_data_label[0], y_data_size, 1, 0)
image_label = sess.run(tf.cast(lbl_one_hot, tf.uint8)).tobytes()
feature = {
'label': _bytes_feature(image_label),
'image': _bytes_feature(image_bytes)
}
example = tf.train.Example(features=tf.train.Features(
feature=feature))
writer.write(example.SerializeToString())
current_index += 1
writer.close()
def batch_crop_image_parts(ori_set, target_set):
for file in os.listdir(ori_set):
print(file)
image_file = os.path.join(ori_set, str(file), str(file) + '.bmp')
det_mask_file = os.path.join(ori_set, str(file),
str(file) + '_detection.bmp')
cls_mask_file = os.path.join(ori_set, str(file),
str(file) + '_classification.bmp')
image = cv2.imread(image_file)
image = cv2.resize(image, (512, 512))
det_mask = cv2.imread(det_mask_file)
det_mask = cv2.resize(det_mask, (512, 512))
cls_mask = cv2.imread(cls_mask_file)
cls_mask = cv2.resize(cls_mask, (512, 512))
crop_list = crop_image_parts(image, det_mask, cls_mask)
list_file_create = [
os.path.join(target_set,
str(file) + '_1'),
os.path.join(target_set,
str(file) + '_2'),
os.path.join(target_set,
str(file) + '_3'),
os.path.join(target_set,
str(file) + '_4')
]
check_directory(list_file_create)
list_img_create = [
os.path.join(target_set,
str(file) + '_1',
str(file) + '_1.bmp'),
os.path.join(target_set,
str(file) + '_2',
str(file) + '_2.bmp'),
os.path.join(target_set,
str(file) + '_3',
str(file) + '_3.bmp'),
os.path.join(target_set,
str(file) + '_4',
str(file) + '_4.bmp'),
os.path.join(target_set,
str(file) + '_1',
str(file) + '_1_detection.bmp'),
os.path.join(target_set,
str(file) + '_2',
str(file) + '_2_detection.bmp'),
os.path.join(target_set,
str(file) + '_3',
str(file) + '_3_detection.bmp'),
os.path.join(target_set,
str(file) + '_4',
str(file) + '_4_detection.bmp'),
os.path.join(target_set,
str(file) + '_1',
str(file) + '_1_classification.bmp'),
os.path.join(target_set,
str(file) + '_2',
str(file) + '_2_classification.bmp'),
os.path.join(target_set,
str(file) + '_3',
str(file) + '_3_classification.bmp'),
os.path.join(target_set,
str(file) + '_4',
str(file) + '_4_classification.bmp'),
os.path.join(target_set,
str(file) + '_1',
str(file) + '_1_original.bmp'),
os.path.join(target_set,
str(file) + '_2',
str(file) + '_2_original.bmp'),
os.path.join(target_set,
str(file) + '_3',
str(file) + '_3_original.bmp'),
os.path.join(target_set,
str(file) + '_4',
str(file) + '_4_original.bmp')
]
for order, img in enumerate(crop_list):
check_cv2_imwrite(list_img_create[order], img)