--There is a method of padding in case there are few learning images. --There are various things such as contrast, gamma, blur, and noise. ――This time, we performed left-right reversal and random cropping. ――Actually, after various experiments, it was the combination with the best learning accuracy. It is the case of the original image this time. --The complete source is here.
--I used Numpy`` Pillow.
$ pip install numpy==1.16.5 pillow
--Sequential processing is repeated according to CLASSES.
--The face image is saved in FACE_PATH.
--According to TEST_NUM, the image is duplicated from FACE_PATH to TEST_PATH.
--TRAIN_PATH duplicates images that were not replicated to TEST_PATH.
--According to ʻAUGMENT_NUM, an inflated image is created from TRAIN_PATH to ʻAUGMENT_PATH.
config.py
CLASSES = [
'Abe Oto',
'Satomi Ishihara',
'Yuno Ohara',
'Fuka Koshiba',
'Haruna Kawaguchi',
'Nana Mori',
'Minami Hamabe',
'Kaya Kiyohara',
'Haruka Fukuhara',
'Kuroshima Yuina'
]
BASE_PATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
DATA_PATH = os.path.join(BASE_PATH, 'data')
FACE_PATH = os.path.join(DATA_PATH, 'face')
TRAIN_PATH = os.path.join(DATA_PATH, 'train')
TEST_PATH = os.path.join(DATA_PATH, 'test')
AUGMENT_PATH = os.path.join(DATA_PATH, 'augment')
TRAIN_NUM = 0
TEST_NUM = 100
AUGMENT_NUM = 6000
--Check the path of the face image, learning image, and test image.
--Create a list of face images.
--query is given sequentially CLASSES.
save_train_test_from_face.py
def split(query):
"""Get a list of face images, divide and copy into learning and testing."""
face_path = os.path.join(FACE_PATH, query)
train_path = os.path.join(TRAIN_PATH, query)
test_path = os.path.join(TEST_PATH, query)
face_file_list = glob.glob(os.path.join(face_path, '*.jpeg'))
face_file_list.sort()
--Shuffle the list of face images.
--According to TEST_NUM, the list of face images is divided into training images and test images.
save_train_test_from_face.py
random.shuffle(face_file_list)
train_file_list = face_file_list[:-TEST_NUM]
test_file_list = face_file_list[len(train_file_list):]
--Create a duplicate of the training image and the test image. ――If you keep the original face image, you can save the trouble of redoing it.
save_train_test_from_face.py
for face_file in train_file_list:
train_file = os.path.join(train_path, os.path.basename(face_file))
shutil.copy(face_file, train_file)
for face_file in test_file_list:
test_file = os.path.join(test_path, os.path.basename(face_file))
shutil.copy(face_file, test_file)
――The face image was divided into a learning image and a test image as shown below.
――The maximum number of learning images is 392 and the minimum number is 269. Maybe few.
$ python save_train_test_from_face.py
query:Abe Oto, face: 415, train: 315, test: 100
query:Satomi Ishihara, face: 492, train: 392, test: 100
query:Yuno Ohara, face: 372, train: 272, test: 100
query:Fuka Koshiba, face: 400, train: 300, test: 100
query:Haruna Kawaguchi, face: 369, train: 269, test: 100
query:Nana Mori, face: 389, train: 289, test: 100
query:Minami Hamabe, face: 481, train: 381, test: 100
query:Kaya Kiyohara, face: 428, train: 328, test: 100
query:Haruka Fukuhara, face: 420, train: 320, test: 100
query:Kuroshima Yuina, face: 448, train: 348, test: 100
――I referred to the following. -Summary of Data Augmentation of images in NumPy
--First, convert from Pillow to Numpy.
--Also, rate gives the probability of inversion. We set 0.5 to a 50-50 chance.
--Convert to Numpy and flip it horizontally with fliplr.
--Finally, change from Numpy back to Pillow.
def horizontal_flip(image, rate=0.5):
"""Flip horizontally."""
image = np.array(image, dtype=np.float32)
if np.random.rand() < rate:
image = np.fliplr(image)
return Image.fromarray(np.uint8(image))
--Get the height and width of the image with ʻimage.shape. --Determine the crop size based on size. 0.8 means to crop at a size of 80% . --Position the upper leftandlower right. --top is a random value in the range 0toheight -- crop_size. --Similarly, determine left. --bottom is located by adding topandcrop_size. --Similarly, determine right. --Finally, crop from ʻimage.
def random_crop(image, size=0.8):
"""Crop at random size."""
image = np.array(image, dtype=np.float32)
height, width, _ = image.shape
crop_size = int(min(height, width) * size)
top = np.random.randint(0, height - crop_size)
left = np.random.randint(0, width - crop_size)
bottom = top + crop_size
right = left + crop_size
image = image[top:bottom, left:right, :]
return Image.fromarray(np.uint8(image))
--Set the path for the training image and the padded image.
--query is given sequentially CLASSES.
def augment(query):
"""Load, inflate, and save learning images."""
train_path = os.path.join(TRAIN_PATH, query)
augment_path = os.path.join(AUGMENT_PATH, query)
--Create a list of face images.
train_list = glob.glob(os.path.join(train_path, '*.jpeg'))
train_list.sort()
--From the number of padded images, check how many face images should be created and determine the number of loop processes.
loop_num = math.ceil(AUGMENT_NUM / len(train_list))
--Perform the following in the loop processing count and face image list loop.
--Reading the face image.
――Inverts horizontally at a rate of 50%.
--Random crop with 80% image size.
--Add -0001.jpeg to the file name of the face image and save the inflated image.
augment_num = 0
for num in range(1, loop_num + 1):
for train_file in train_list:
if augment_num == AUGMENT_NUM:
break
image = Image.open(train_file)
image = horizontal_flip(image)
image = random_crop(image)
augment_file = os.path.join(AUGMENT_PATH, query, os.path.basename(train_file).split('.')[0] + '-{:04d}.jpeg'.format(num))
image.save(augment_file, optimize=True, quality=95)
print('query: {}, train_file: {}, augment_file: {}'.format(
query, os.path.basename(train_file), os.path.basename(augment_file)))
augment_num += 1
--The training image was inflated with Pillo and Numpy.
――In the process of work, we also confirmed scale crop, cut apto, random erase, and random rotate other than random crop. In the case of the face image this time, I did not use it because it was not suitable for improving accuracy.
――Next time, we plan to create a data set to make the training images and test images easier to handle.
Recommended Posts