def _get_features(self, frame, cell_label):
"""Gets the features of the cell in the frame.
Args:
frame (int): frame from which to get cell features.
cell_label (int): label of the cell.
Returns:
dict: a dictionary with keys as the feature names.
"""
# Get the bounding box
X_frame = self._get_frame(self.x, frame)
y_frame = self._get_frame(self.y, frame)
roi = (y_frame == cell_label).astype('int32')
props = regionprops(np.squeeze(roi), coordinates='rc')[0]
centroid = props.centroid
rprop = np.array([
props.area,
props.perimeter,
props.eccentricity
])
# Extract images from bounding boxes
minr, minc, maxr, maxc = props.bbox
if self.data_format == 'channels_first':
appearance = np.copy(X_frame[:, minr:maxr, minc:maxc])
else:
appearance = np.copy(X_frame[minr:maxr, minc:maxc, :])
# Resize images from bounding box
appearance = resize(appearance, (self.crop_dim, self.crop_dim),
data_format=self.data_format)
# Get the neighborhood
neighborhood = self._sub_area(X_frame, y_frame, cell_label)
# Try to assign future areas if future frame is available
# TODO: We shouldn't grab a future frame if the frame is dark (was padded)
try:
X_future_frame = self._get_frame(self.x, frame + 1)
future_area = self._sub_area(X_future_frame, y_frame, cell_label)
except IndexError:
future_area = neighborhood
# future areas are not a feature instead a part of the neighborhood feature
return {
'appearance': np.expand_dims(appearance, axis=0),
'distance': np.expand_dims(centroid, axis=0),
'neighborhood': np.expand_dims(neighborhood, axis=0),
'regionprop': np.expand_dims(rprop, axis=0),
'~future area': np.expand_dims(future_area, axis=0)
}
def test_resize(self):
channel_sizes = (3, 1) # skimage used for multi-channel, cv2 otherwise
for c in channel_sizes:
for data_format in ('channels_last', 'channels_first'):
channel_axis = 2 if data_format == 'channels_last' else 0
img = np.stack([_get_image()] * c, axis=channel_axis)
resize_shape = (28, 28)
resized_img = utils.resize(img, resize_shape,
data_format=data_format)
if data_format == 'channels_first':
assert resized_img.shape[1:] == resize_shape
else:
assert resized_img.shape[:-1] == resize_shape
def _sub_area(self, X_frame, y_frame, cell_label):
"""Fetch a neighborhood surrounding the cell in the given frame.
Slices out a neighborhood_true_size square region around the center of
the provided cell, and reshapes it to neighborhood_scale_size square.
Args:
X_frame (np.array): 2D numpy array, a frame of raw data.
y_frame (np.array): 2D numpy array, a frame of annotated data.
cell_label (int): The label of the cell to slice out.
Returns:
numpy.array: the resized region of X_frame around cell_label.
"""
pads = ((self.neighborhood_true_size, self.neighborhood_true_size),
(self.neighborhood_true_size,
self.neighborhood_true_size), (0, 0))
X_padded = np.pad(X_frame, pads, mode='constant', constant_values=0)
y_padded = np.pad(y_frame, pads, mode='constant', constant_values=0)
roi = (y_padded == cell_label).astype('int32')
props = regionprops(np.squeeze(roi), coordinates='rc')
center_x, center_y = props[0].centroid
center_x, center_y = np.int(center_x), np.int(center_y)
x1 = center_x - self.neighborhood_true_size
x2 = center_x + self.neighborhood_true_size
y1 = center_y - self.neighborhood_true_size
y2 = center_y + self.neighborhood_true_size
X_reduced = X_padded[x1:x2, y1:y2]
# resize to neighborhood_scale_size
resize_shape = (2 * self.neighborhood_scale_size + 1,
2 * self.neighborhood_scale_size + 1)
X_reduced = resize(X_reduced,
resize_shape,
data_format=self.data_format)
return X_reduced