def __init__(self):

self.labels = list()

self.hashes = dict()

self.hash_size = 16

self.data_shape = None

def get_images(self, directory, train=False):

data, label = list(), list()

for filename in os.listdir(directory):

if os.path.isdir(directory + filename) and 'no' in filename:

for fname in os.listdir(directory + filename):

img = cv2.imread(directory + filename + '/' + fname)

data.append(img)

label.append(0)

if os.path.isdir(directory + filename) and 'yes' in filename:

for fname in os.listdir(directory + filename):

img = cv2.imread(directory + filename + '/' + fname)

data.append(img)

label.append(1)

if train:

return self._data_augmentation(data, label)

return np.array(data), np.array(label)

def _data_augmentation(self, data, labels):

ddata, llabels = list(), list()

for img, l in zip(data, labels):

# Rotate images by 90 degrees

ddata.append(cv2.rotate(img, cv2.ROTATE_90_CLOCKWISE))

llabels.append(l)

for img, l in zip(data, labels):

# Flip images vertically

ddata.append(cv2.flip(img, 0))

llabels.append(l)

return np.concatenate((np.array(data), np.array(ddata)), axis=None), np.concatenate((np.array(labels), np.array(llabels)), axis=None)

def check_duplicates(self, fname):

with Image.open(fname) as img:

temp_hash = imagehash.average_hash(img, self.hash_size)

if temp_hash in self.hashes:

return True

else:

self.hashes[temp_hash] = fname

def _aug_transform(self, data, label):

augs = image.CreateAugmenter(data_shape=self.data_shape,

rand_crop=0.5,

rand_mirror=True,

inter_method=10,

brightness=0.125, contrast=0.125, saturation=0.125, pca_noise=0.02)

for aug in augs:

data = aug(data)

return data, label

def data_loader(self, data, label, data_shape, transform=True, shuffle=True, batch_size=32):

self.data_shape = data_shape

for i, d in enumerate(data):

d = d.astype('float32') / 255

if transform:

data[i], label = self._aug_transform(nd.array(d), label)

else:

data[i]= image.imresize(nd.array(data[i]), w=data_shape[1], h=data_shape[2])

data[i] = nd.transpose(data=nd.array(data[i]), axes=(2, 0, 1))

dataset = gluon.data.dataset.ArrayDataset(data, label)

dataloader = gluon.data.DataLoader(dataset, batch_size=batch_size, shuffle=shuffle)

return dataloader

def crop_imgs(self, imgs, add_pixels_value = 0):

new_images = list()

for img in imgs:

gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)

gray = cv2.GaussianBlur(gray, (5, 5), 0)

# Threshold the image then perform erosions and dilations to remove

# small regions of noise

_, thresh = cv2.threshold(gray, 45, 255, cv2.THRESH_BINARY)

thresh = cv2.erode(thresh, None, iterations=2)

thresh = cv2.dilate(thresh, None, iterations=2)

# Find contours in threshold image

contours = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

contours = imutils.grab_contours(contours)

c = max(contours, key=cv2.contourArea)

extLeft = (c[c[:, :, 0].argmin()][0])

extRight = c[c[:, :, 0].argmax()][0]

extTop = c[c[:, :, 1].argmin()][0]

extBottom= c[c[:, :, 1].argmax()][0]

new_img = img[extTop[1] - add_pixels_value : extBottom[1] + add_pixels_value, \

extLeft[0] - add_pixels_value : extRight[0] + add_pixels_value].copy()

new_images.append(new_img)

return np.array(new_images, dtype='object')

def save_images(self, images, labels, directory):

file_names = ['cropped{}.jpg'.format(i) for i in range(len(images))]

for i, image in enumerate(images):

if labels[i] == 1:

d = directory + 'yes/'

cv2.imwrite(d + file_names[i], image)

else:

d = directory + 'no/'