def __getitem__(
self, index: int
) -> Tuple[np.ndarray, np.ndarray, np.ndarray, torch.Tensor]:
kind, file, label = (
self.kinds[index],
self.files[index],
self.labels[index],
)
dct_y, dct_cb, dct_cr = dct_from_jpeg_imageio(file)
dct_y = dct_y.astype(np.float32)
dct_cb = dct_cb.astype(np.float32)
dct_cr = dct_cr.astype(np.float32)
dct_y = np.rollaxis(dct_y, 2, 0)
dct_cb = np.rollaxis(dct_cb, 2, 0)
dct_cr = np.rollaxis(dct_cr, 2, 0)
dct_y = dct_y / 1024
dct_cb = dct_cb / 1024
dct_cr = dct_cr / 1024
target = one_hot(10, label)
return dct_y, dct_cb, dct_cr, target
def __getitem__(
self, index: int
) -> Tuple[np.ndarray, np.ndarray, np.ndarray, torch.Tensor]:
kind, image_name, label = (
self.kinds[index],
self.image_names[index],
self.labels[index],
)
dct_y, dct_cb, dct_cr = dct_from_jpeg_imageio(
f"data/{kind}/{image_name}"
)
dct_y = dct_y.astype(np.float32)
dct_cb = dct_cb.astype(np.float32)
dct_cr = dct_cr.astype(np.float32)
dct_y = np.rollaxis(dct_y, 2, 0)
dct_cb = np.rollaxis(dct_cb, 2, 0)
dct_cr = np.rollaxis(dct_cr, 2, 0)
dct_y = dct_y / 1024
dct_cb = dct_cb / 1024
dct_cr = dct_cr / 1024
target = one_hot(self.n_classes, label)
return dct_y, dct_cb, dct_cr, target
def __getitem__(self, index):
kind, image_name, label = (
self.kinds[index],
self.image_names[index],
self.labels[index],
)
dct_y, dct_cb, dct_cr = dct_from_jpeg_imageio(
f"data/{kind}/{image_name}")
dct_y = dct_y.astype(np.float32)
dct_cb = dct_cb.astype(np.float32)
dct_cr = dct_cr.astype(np.float32)
# dct_y = np.rollaxis(dct_y, 2, 0)
# dct_cb = np.rollaxis(dct_cb, 2, 0)
# dct_cr = np.rollaxis(dct_cr, 2, 0)
dct_y = dct_y / 1024
dct_cb = dct_cb / 1024
dct_cr = dct_cr / 1024
# Flatten each array from shape (64, 64, 64) to (4096, 64)
dct_y = dct_y.reshape((4096, 64))
dct_cb = dct_cb.reshape((4096, 64))
dct_cr = dct_cr.reshape((4096, 64))
# Concatenate the arrays
input_data = np.concatenate((dct_y, dct_cb, dct_cr), axis=0)
target = one_hot(4, label)
return input_data, target
def pre_process_and_save_dct_data():
for kind in tqdm(["Cover", "JMiPOD", "JUNIWARD", "UERD"], desc=""):
for file in tqdm(glob.glob1(f"data/Cover/", "*.jpg"), desc=""):
save_dir = f"data/dct/{kind}/"
load_dir = f"data/{kind}/"
save_file = f"{save_dir}/{file}".replace(".jpg", ".npz")
load_file = f"{load_dir}/{file}"
make_dir_if_not_exists(save_dir)
# Load the DCT arrays
dct_y, dct_cb, dct_cr = dct_from_jpeg_imageio(load_file)
# Save the arrays without any normalisation
np.savez_compressed(save_file, dct_y, dct_cb, dct_cr)
def test_dct_methods() -> None:
test_path = "data/UERD/00001.jpg"
uber_dct_y, uber_dct_cb, uber_dct_cr = dct_from_jpeg(test_path)
imageio_dct_y, imageio_dct_cb, imageio_dct_cr = dct_from_jpeg_imageio(
test_path
)
assert uber_dct_y.shape == (64, 64, 64)
assert uber_dct_cb.shape == (32, 32, 64)
assert uber_dct_cr.shape == (32, 32, 64)
assert imageio_dct_y.shape == (64, 64, 64)
assert imageio_dct_cb.shape == (64, 64, 64)
assert imageio_dct_cr.shape == (64, 64, 64)
def __getitem__(self, index):
image_name = self.image_names[index]
img_path = f"data/Test/{image_name}"
dct_y, dct_cb, dct_cr = dct_from_jpeg_imageio(img_path)
dct_y = dct_y.astype(np.float32)
dct_cb = dct_cb.astype(np.float32)
dct_cr = dct_cr.astype(np.float32)
dct_y = np.rollaxis(dct_y, 2, 0)
dct_cb = np.rollaxis(dct_cb, 2, 0)
dct_cr = np.rollaxis(dct_cr, 2, 0)
dct_y = dct_y / 1024
dct_cb = dct_cb / 1024
dct_cr = dct_cr / 1024
return image_name, [dct_y, dct_cb, dct_cr]
def load_dct_values_from_pre_processed_imagenet_image(image_path,
is_training=True):
# Read the image with CV2 and convert to BGR
img = cv2.imread(image_path, cv2.IMREAD_UNCHANGED)
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
if is_training:
try:
img = get_random_crop(img, image_size=IMAGE_SIZE)
except ValueError:
print("Invalid random crop, reverting to center crop.")
img = get_center_crop(img, image_size=IMAGE_SIZE)
else:
img = get_center_crop(img, image_size=IMAGE_SIZE)
with tempfile.TemporaryDirectory() as tmp_dir:
tmp_file = f"{tmp_dir}/img.jpg"
# Convert to PIL and save without subsampling
img = Image.fromarray(np.array(img))
img.save(tmp_file, subsampling=0, format="JPEG")
return dct_from_jpeg_imageio(tmp_file)
def compare_cover_vs_modified_dct() -> None:
dct_y_0, dct_cb_0, dct_cr_0 = dct_from_jpeg_imageio("data/UERD/00001.jpg")
dct_y_1, dct_cb_1, dct_cr_1 = dct_from_jpeg_imageio("data/UERD/00001.jpg")
print(dct_y_0[0])
print(dct_y_1[0])