forked from 170010011/fr
161 lines
5.4 KiB
Python
161 lines
5.4 KiB
Python
from skimage.feature import multiscale_basic_features
|
|
|
|
try:
|
|
from sklearn.exceptions import NotFittedError
|
|
from sklearn.ensemble import RandomForestClassifier
|
|
has_sklearn = True
|
|
except ImportError:
|
|
has_sklearn = False
|
|
|
|
class NotFittedError(Exception):
|
|
pass
|
|
|
|
|
|
class TrainableSegmenter(object):
|
|
"""Estimator for classifying pixels.
|
|
|
|
Parameters
|
|
----------
|
|
clf : classifier object, optional
|
|
classifier object, exposing a ``fit`` and a ``predict`` method as in
|
|
scikit-learn's API, for example an instance of
|
|
``RandomForestClassifier`` or ``LogisticRegression`` classifier.
|
|
features_func : function, optional
|
|
function computing features on all pixels of the image, to be passed
|
|
to the classifier. The output should be of shape
|
|
``(m_features, *labels.shape)``. If None,
|
|
:func:`skimage.segmentation.multiscale_basic_features` is used.
|
|
|
|
Methods
|
|
-------
|
|
compute_features
|
|
fit
|
|
predict
|
|
"""
|
|
|
|
def __init__(self, clf=None, features_func=None):
|
|
if clf is None:
|
|
if has_sklearn:
|
|
self.clf = RandomForestClassifier(n_estimators=100, n_jobs=-1)
|
|
else:
|
|
raise ImportError(
|
|
"Please install scikit-learn or pass a classifier instance"
|
|
"to TrainableSegmenter."
|
|
)
|
|
else:
|
|
self.clf = clf
|
|
self.features_func = features_func
|
|
|
|
def compute_features(self, image):
|
|
if self.features_func is None:
|
|
self.features_func = multiscale_basic_features
|
|
self.features = self.features_func(image)
|
|
|
|
def fit(self, image, labels):
|
|
"""Train classifier using partially labeled (annotated) image.
|
|
|
|
Parameters
|
|
----------
|
|
image : ndarray
|
|
Input image, which can be grayscale or multichannel, and must have a
|
|
number of dimensions compatible with ``self.features_func``.
|
|
labels : ndarray of ints
|
|
Labeled array of shape compatible with ``image`` (same shape for a
|
|
single-channel image). Labels >= 1 correspond to the training set and
|
|
label 0 to unlabeled pixels to be segmented.
|
|
"""
|
|
self.compute_features(image)
|
|
clf = fit_segmenter(labels, self.features, self.clf)
|
|
|
|
def predict(self, image):
|
|
"""Segment new image using trained internal classifier.
|
|
|
|
Parameters
|
|
----------
|
|
image : ndarray
|
|
Input image, which can be grayscale or multichannel, and must have a
|
|
number of dimensions compatible with ``self.features_func``.
|
|
|
|
Raises
|
|
------
|
|
NotFittedError if ``self.clf`` has not been fitted yet (use ``self.fit``).
|
|
"""
|
|
if self.features_func is None:
|
|
self.features_func = multiscale_basic_features
|
|
features = self.features_func(image)
|
|
return predict_segmenter(features, self.clf)
|
|
|
|
|
|
def fit_segmenter(labels, features, clf):
|
|
"""Segmentation using labeled parts of the image and a classifier.
|
|
|
|
Parameters
|
|
----------
|
|
labels : ndarray of ints
|
|
Image of labels. Labels >= 1 correspond to the training set and
|
|
label 0 to unlabeled pixels to be segmented.
|
|
features : ndarray
|
|
Array of features, with the first dimension corresponding to the number
|
|
of features, and the other dimensions correspond to ``labels.shape``.
|
|
clf : classifier object
|
|
classifier object, exposing a ``fit`` and a ``predict`` method as in
|
|
scikit-learn's API, for example an instance of
|
|
``RandomForestClassifier`` or ``LogisticRegression`` classifier.
|
|
|
|
Returns
|
|
-------
|
|
clf : classifier object
|
|
classifier trained on ``labels``
|
|
|
|
Raises
|
|
------
|
|
NotFittedError if ``self.clf`` has not been fitted yet (use ``self.fit``).
|
|
"""
|
|
mask = labels > 0
|
|
training_data = features[mask]
|
|
training_labels = labels[mask].ravel()
|
|
clf.fit(training_data, training_labels)
|
|
return clf
|
|
|
|
|
|
def predict_segmenter(features, clf):
|
|
"""Segmentation of images using a pretrained classifier.
|
|
|
|
Parameters
|
|
----------
|
|
features : ndarray
|
|
Array of features, with the last dimension corresponding to the number
|
|
of features, and the other dimensions are compatible with the shape of
|
|
the image to segment, or a flattened image.
|
|
clf : classifier object
|
|
trained classifier object, exposing a ``predict`` method as in
|
|
scikit-learn's API, for example an instance of
|
|
``RandomForestClassifier`` or ``LogisticRegression`` classifier. The
|
|
classifier must be already trained, for example with
|
|
:func:`skimage.segmentation.fit_segmenter`.
|
|
|
|
Returns
|
|
-------
|
|
output : ndarray
|
|
Labeled array, built from the prediction of the classifier.
|
|
"""
|
|
sh = features.shape
|
|
if features.ndim > 2:
|
|
features = features.reshape((-1, sh[-1]))
|
|
|
|
try:
|
|
predicted_labels = clf.predict(features)
|
|
except NotFittedError:
|
|
raise NotFittedError(
|
|
"You must train the classifier `clf` first"
|
|
"for example with the `fit_segmenter` function."
|
|
)
|
|
except ValueError as err:
|
|
if err.args and 'x must consist of vectors of length' in err.args[0]:
|
|
raise ValueError(
|
|
err.args[0] + '\n' +
|
|
"Maybe you did not use the same type of features for training the classifier."
|
|
)
|
|
output = predicted_labels.reshape(sh[:-1])
|
|
return output
|