image.py

"""
The :mod:`sklearn.feature_extraction.image` submodule gathers utilities to
extract features from images.
"""

# Authors: Emmanuelle Gouillart <emmanuelle.gouillart@normalesup.org>
#          Gael Varoquaux <gael.varoquaux@normalesup.org>
#          Olivier Grisel
#          Vlad Niculae
# License: BSD 3 clause

from itertools import product
from numbers import Integral, Number, Real

import numpy as np
from numpy.lib.stride_tricks import as_strided
from scipy import sparse

from ..base import BaseEstimator, TransformerMixin, _fit_context
from ..utils import check_array, check_random_state
from ..utils._param_validation import Hidden, Interval, RealNotInt, validate_params

__all__ = [
    "PatchExtractor",
    "extract_patches_2d",
    "grid_to_graph",
    "img_to_graph",
    "reconstruct_from_patches_2d",
]

###############################################################################
# From an image to a graph


def _make_edges_3d(n_x, n_y, n_z=1):
    """Returns a list of edges for a 3D image.

    Parameters
    ----------
    n_x : int
        The size of the grid in the x direction.
    n_y : int
        The size of the grid in the y direction.
    n_z : integer, default=1
        The size of the grid in the z direction, defaults to 1
    """
    vertices = np.arange(n_x * n_y * n_z).reshape((n_x, n_y, n_z))
    edges_deep = np.vstack((vertices[:, :, :-1].ravel(), vertices[:, :, 1:].ravel()))
    edges_right = np.vstack((vertices[:, :-1].ravel(), vertices[:, 1:].ravel()))
    edges_down = np.vstack((vertices[:-1].ravel(), vertices[1:].ravel()))
    edges = np.hstack((edges_deep, edges_right, edges_down))
    return edges


def _compute_gradient_3d(edges, img):
    _, n_y, n_z = img.shape
    gradient = np.abs(
        img[
            edges[0] // (n_y * n_z),
            (edges[0] % (n_y * n_z)) // n_z,
            (edges[0] % (n_y * n_z)) % n_z,
        ]
        - img[
            edges[1] // (n_y * n_z),
            (edges[1] % (n_y * n_z)) // n_z,
            (edges[1] % (n_y * n_z)) % n_z,
        ]
    )
    return gradient


# XXX: Why mask the image after computing the weights?


def _mask_edges_weights(mask, edges, weights=None):
    """Apply a mask to edges (weighted or not)"""
    inds = np.arange(mask.size)
    inds = inds[mask.ravel()]
    ind_mask = np.logical_and(np.isin(edges[0], inds), np.isin(edges[1], inds))
    edges = edges[:, ind_mask]
    if weights is not None:
        weights = weights[ind_mask]
    if len(edges.ravel()):
        maxval = edges.max()
    else:
        maxval = 0
    order = np.searchsorted(np.flatnonzero(mask), np.arange(maxval + 1))
    edges = order[edges]
    if weights is None:
        return edges
    else:
        return edges, weights


def _to_graph(
    n_x, n_y, n_z, mask=None, img=None, return_as=sparse.coo_matrix, dtype=None
):
    """Auxiliary function for img_to_graph and grid_to_graph"""
    edges = _make_edges_3d(n_x, n_y, n_z)

    if dtype is None:  # To not overwrite input dtype
        if img is None:
            dtype = int
        else:
            dtype = img.dtype

    if img is not None:
        img = np.atleast_3d(img)
        weights = _compute_gradient_3d(edges, img)
        if mask is not None:
            edges, weights = _mask_edges_weights(mask, edges, weights)
            diag = img.squeeze()[mask]
        else:
            diag = img.ravel()
        n_voxels = diag.size
    else:
        if mask is not None:
            mask = mask.astype(dtype=bool, copy=False)
            edges = _mask_edges_weights(mask, edges)
            n_voxels = np.sum(mask)
        else:
            n_voxels = n_x * n_y * n_z
        weights = np.ones(edges.shape[1], dtype=dtype)
        diag = np.ones(n_voxels, dtype=dtype)

    diag_idx = np.arange(n_voxels)
    i_idx = np.hstack((edges[0], edges[1]))
    j_idx = np.hstack((edges[1], edges[0]))
    graph = sparse.coo_matrix(
        (
            np.hstack((weights, weights, diag)),
            (np.hstack((i_idx, diag_idx)), np.hstack((j_idx, diag_idx))),
        ),
        (n_voxels, n_voxels),
        dtype=dtype,
    )
    if return_as is np.ndarray:
        return graph.toarray()
    return return_as(graph)


@validate_params(
    {
        "img": ["array-like"],
        "mask": [None, np.ndarray],
        "return_as": [type],
        "dtype": "no_validation",  # validation delegated to numpy
    },
    prefer_skip_nested_validation=True,
)
def img_to_graph(img, *, mask=None, return_as=sparse.coo_matrix, dtype=None):
    """Graph of the pixel-to-pixel gradient connections.

    Edges are weighted with the gradient values.

    Read more in the :ref:`User Guide <image_feature_extraction>`.

    Parameters
    ----------
    img : array-like of shape (height, width) or (height, width, channel)
        2D or 3D image.
    mask : ndarray of shape (height, width) or \
            (height, width, channel), dtype=bool, default=None
        An optional mask of the image, to consider only part of the
        pixels.
    return_as : np.ndarray or a sparse matrix class, \
            default=sparse.coo_matrix
        The class to use to build the returned adjacency matrix.
    dtype : dtype, default=None
        The data of the returned sparse matrix. By default it is the
        dtype of img.

    Returns
    -------
    graph : ndarray or a sparse matrix class
        The computed adjacency matrix.

    Notes
    -----
    For scikit-learn versions 0.14.1 and prior, return_as=np.ndarray was
    handled by returning a dense np.matrix instance.  Going forward, np.ndarray
    returns an np.ndarray, as expected.

    For compatibility, user code relying on this method should wrap its
    calls in ``np.asarray`` to avoid type issues.
    """
    img = np.atleast_3d(img)
    n_x, n_y, n_z = img.shape
    return _to_graph(n_x, n_y, n_z, mask, img, return_as, dtype)


@validate_params(
    {
        "n_x": [Interval(Integral, left=1, right=None, closed="left")],
        "n_y": [Interval(Integral, left=1, right=None, closed="left")],
        "n_z": [Interval(Integral, left=1, right=None, closed="left")],
        "mask": [None, np.ndarray],
        "return_as": [type],
        "dtype": "no_validation",  # validation delegated to numpy
    },
    prefer_skip_nested_validation=True,
)
def grid_to_graph(
    n_x, n_y, n_z=1, *, mask=None, return_as=sparse.coo_matrix, dtype=int
):
    """Graph of the pixel-to-pixel connections.

    Edges exist if 2 voxels are connected.

    Parameters
    ----------
    n_x : int
        Dimension in x axis.
    n_y : int
        Dimension in y axis.
    n_z : int, default=1
        Dimension in z axis.
    mask : ndarray of shape (n_x, n_y, n_z), dtype=bool, default=None
        An optional mask of the image, to consider only part of the
        pixels.
    return_as : np.ndarray or a sparse matrix class, \
            default=sparse.coo_matrix
        The class to use to build the returned adjacency matrix.
    dtype : dtype, default=int
        The data of the returned sparse matrix. By default it is int.

    Returns
    -------
    graph : np.ndarray or a sparse matrix class
        The computed adjacency matrix.

    Notes
    -----
    For scikit-learn versions 0.14.1 and prior, return_as=np.ndarray was
    handled by returning a dense np.matrix instance.  Going forward, np.ndarray
    returns an np.ndarray, as expected.

    For compatibility, user code relying on this method should wrap its
    calls in ``np.asarray`` to avoid type issues.
    """
    return _to_graph(n_x, n_y, n_z, mask=mask, return_as=return_as, dtype=dtype)


###############################################################################
# From an image to a set of small image patches


def _compute_n_patches(i_h, i_w, p_h, p_w, max_patches=None):
    """Compute the number of patches that will be extracted in an image.

    Read more in the :ref:`User Guide <image_feature_extraction>`.

    Parameters
    ----------
    i_h : int
        The image height
    i_w : int
        The image with
    p_h : int
        The height of a patch
    p_w : int
        The width of a patch
    max_patches : int or float, default=None
        The maximum number of patches to extract. If `max_patches` is a float
        between 0 and 1, it is taken to be a proportion of the total number
        of patches. If `max_patches` is None, all possible patches are extracted.
    """
    n_h = i_h - p_h + 1
    n_w = i_w - p_w + 1
    all_patches = n_h * n_w

    if max_patches:
        if isinstance(max_patches, (Integral)) and max_patches < all_patches:
            return max_patches
        elif isinstance(max_patches, (Integral)) and max_patches >= all_patches:
            return all_patches
        elif isinstance(max_patches, (Real)) and 0 < max_patches < 1:
            return int(max_patches * all_patches)
        else:
            raise ValueError("Invalid value for max_patches: %r" % max_patches)
    else:
        return all_patches


def _extract_patches(arr, patch_shape=8, extraction_step=1):
    """Extracts patches of any n-dimensional array in place using strides.

    Given an n-dimensional array it will return a 2n-dimensional array with
    the first n dimensions indexing patch position and the last n indexing
    the patch content. This operation is immediate (O(1)). A reshape
    performed on the first n dimensions will cause numpy to copy data, leading
    to a list of extracted patches.

    Read more in the :ref:`User Guide <image_feature_extraction>`.

    Parameters
    ----------
    arr : ndarray
        n-dimensional array of which patches are to be extracted

    patch_shape : int or tuple of length arr.ndim.default=8
        Indicates the shape of the patches to be extracted. If an
        integer is given, the shape will be a hypercube of
        sidelength given by its value.

    extraction_step : int or tuple of length arr.ndim, default=1
        Indicates step size at which extraction shall be performed.
        If integer is given, then the step is uniform in all dimensions.


    Returns
    -------
    patches : strided ndarray
        2n-dimensional array indexing patches on first n dimensions and
        containing patches on the last n dimensions. These dimensions
        are fake, but this way no data is copied. A simple reshape invokes
        a copying operation to obtain a list of patches:
        result.reshape([-1] + list(patch_shape))
    """

    arr_ndim = arr.ndim

    if isinstance(patch_shape, Number):
        patch_shape = tuple([patch_shape] * arr_ndim)
    if isinstance(extraction_step, Number):
        extraction_step = tuple([extraction_step] * arr_ndim)

    patch_strides = arr.strides

    slices = tuple(slice(None, None, st) for st in extraction_step)
    indexing_strides = arr[slices].strides

    patch_indices_shape = (
        (np.array(arr.shape) - np.array(patch_shape)) // np.array(extraction_step)
    ) + 1

    shape = tuple(list(patch_indices_shape) + list(patch_shape))
    strides = tuple(list(indexing_strides) + list(patch_strides))

    patches = as_strided(arr, shape=shape, strides=strides)
    return patches


@validate_params(
    {
        "image": [np.ndarray],
        "patch_size": [tuple, list],
        "max_patches": [
            Interval(RealNotInt, 0, 1, closed="neither"),
            Interval(Integral, 1, None, closed="left"),
            None,
        ],
        "random_state": ["random_state"],
    },
    prefer_skip_nested_validation=True,
)
def extract_patches_2d(image, patch_size, *, max_patches=None, random_state=None):
    """Reshape a 2D image into a collection of patches.

    The resulting patches are allocated in a dedicated array.

    Read more in the :ref:`User Guide <image_feature_extraction>`.

    Parameters
    ----------
    image : ndarray of shape (image_height, image_width) or \
        (image_height, image_width, n_channels)
        The original image data. For color images, the last dimension specifies
        the channel: a RGB image would have `n_channels=3`.

    patch_size : tuple of int (patch_height, patch_width)
        The dimensions of one patch.

    max_patches : int or float, default=None
        The maximum number of patches to extract. If `max_patches` is a float
        between 0 and 1, it is taken to be a proportion of the total number
        of patches. If `max_patches` is None it corresponds to the total number
        of patches that can be extracted.

    random_state : int, RandomState instance, default=None
        Determines the random number generator used for random sampling when
        `max_patches` is not None. Use an int to make the randomness
        deterministic.
        See :term:`Glossary <random_state>`.

    Returns
    -------
    patches : array of shape (n_patches, patch_height, patch_width) or \
        (n_patches, patch_height, patch_width, n_channels)
        The collection of patches extracted from the image, where `n_patches`
        is either `max_patches` or the total number of patches that can be
        extracted.

    Examples
    --------
    >>> from sklearn.datasets import load_sample_image
    >>> from sklearn.feature_extraction import image
    >>> # Use the array data from the first image in this dataset:
    >>> one_image = load_sample_image("china.jpg")
    >>> print('Image shape: {}'.format(one_image.shape))
    Image shape: (427, 640, 3)
    >>> patches = image.extract_patches_2d(one_image, (2, 2))
    >>> print('Patches shape: {}'.format(patches.shape))
    Patches shape: (272214, 2, 2, 3)
    >>> # Here are just two of these patches:
    >>> print(patches[1])
    [[[174 201 231]
      [174 201 231]]
     [[173 200 230]
      [173 200 230]]]
    >>> print(patches[800])
    [[[187 214 243]
      [188 215 244]]
     [[187 214 243]
      [188 215 244]]]
    """
    i_h, i_w = image.shape[:2]
    p_h, p_w = patch_size

    if p_h > i_h:
        raise ValueError(
            "Height of the patch should be less than the height of the image."
        )

    if p_w > i_w:
        raise ValueError(
            "Width of the patch should be less than the width of the image."
        )

    image = check_array(image, allow_nd=True)
    image = image.reshape((i_h, i_w, -1))
    n_colors = image.shape[-1]

    extracted_patches = _extract_patches(
        image, patch_shape=(p_h, p_w, n_colors), extraction_step=1
    )

    n_patches = _compute_n_patches(i_h, i_w, p_h, p_w, max_patches)
    if max_patches:
        rng = check_random_state(random_state)
        i_s = rng.randint(i_h - p_h + 1, size=n_patches)
        j_s = rng.randint(i_w - p_w + 1, size=n_patches)
        patches = extracted_patches[i_s, j_s, 0]
    else:
        patches = extracted_patches

    patches = patches.reshape(-1, p_h, p_w, n_colors)
    # remove the color dimension if useless
    if patches.shape[-1] == 1:
        return patches.reshape((n_patches, p_h, p_w))
    else:
        return patches


@validate_params(
    {"patches": [np.ndarray], "image_size": [tuple, Hidden(list)]},
    prefer_skip_nested_validation=True,
)
def reconstruct_from_patches_2d(patches, image_size):
    """Reconstruct the image from all of its patches.

    Patches are assumed to overlap and the image is constructed by filling in
    the patches from left to right, top to bottom, averaging the overlapping
    regions.

    Read more in the :ref:`User Guide <image_feature_extraction>`.

    Parameters
    ----------
    patches : ndarray of shape (n_patches, patch_height, patch_width) or \
        (n_patches, patch_height, patch_width, n_channels)
        The complete set of patches. If the patches contain colour information,
        channels are indexed along the last dimension: RGB patches would
        have `n_channels=3`.

    image_size : tuple of int (image_height, image_width) or \
        (image_height, image_width, n_channels)
        The size of the image that will be reconstructed.

    Returns
    -------
    image : ndarray of shape image_size
        The reconstructed image.
    """
    i_h, i_w = image_size[:2]
    p_h, p_w = patches.shape[1:3]
    img = np.zeros(image_size)
    # compute the dimensions of the patches array
    n_h = i_h - p_h + 1
    n_w = i_w - p_w + 1
    for p, (i, j) in zip(patches, product(range(n_h), range(n_w))):
        img[i : i + p_h, j : j + p_w] += p

    for i in range(i_h):
        for j in range(i_w):
            # divide by the amount of overlap
            # XXX: is this the most efficient way? memory-wise yes, cpu wise?
            img[i, j] /= float(min(i + 1, p_h, i_h - i) * min(j + 1, p_w, i_w - j))
    return img


class PatchExtractor(TransformerMixin, BaseEstimator):
    """Extracts patches from a collection of images.

    Read more in the :ref:`User Guide <image_feature_extraction>`.

    .. versionadded:: 0.9

    Parameters
    ----------
    patch_size : tuple of int (patch_height, patch_width), default=None
        The dimensions of one patch. If set to None, the patch size will be
        automatically set to `(img_height // 10, img_width // 10)`, where
        `img_height` and `img_width` are the dimensions of the input images.

    max_patches : int or float, default=None
        The maximum number of patches per image to extract. If `max_patches` is
        a float in (0, 1), it is taken to mean a proportion of the total number
        of patches. If set to None, extract all possible patches.

    random_state : int, RandomState instance, default=None
        Determines the random number generator used for random sampling when
        `max_patches is not None`. Use an int to make the randomness
        deterministic.
        See :term:`Glossary <random_state>`.

    See Also
    --------
    reconstruct_from_patches_2d : Reconstruct image from all of its patches.

    Notes
    -----
    This estimator is stateless and does not need to be fitted. However, we
    recommend to call :meth:`fit_transform` instead of :meth:`transform`, as
    parameter validation is only performed in :meth:`fit`.

    Examples
    --------
    >>> from sklearn.datasets import load_sample_images
    >>> from sklearn.feature_extraction import image
    >>> # Use the array data from the second image in this dataset:
    >>> X = load_sample_images().images[1]
    >>> X = X[None, ...]
    >>> print(f"Image shape: {X.shape}")
    Image shape: (1, 427, 640, 3)
    >>> pe = image.PatchExtractor(patch_size=(10, 10))
    >>> pe_trans = pe.transform(X)
    >>> print(f"Patches shape: {pe_trans.shape}")
    Patches shape: (263758, 10, 10, 3)
    >>> X_reconstructed = image.reconstruct_from_patches_2d(pe_trans, X.shape[1:])
    >>> print(f"Reconstructed shape: {X_reconstructed.shape}")
    Reconstructed shape: (427, 640, 3)
    """

    _parameter_constraints: dict = {
        "patch_size": [tuple, None],
        "max_patches": [
            None,
            Interval(RealNotInt, 0, 1, closed="neither"),
            Interval(Integral, 1, None, closed="left"),
        ],
        "random_state": ["random_state"],
    }

    def __init__(self, *, patch_size=None, max_patches=None, random_state=None):
        self.patch_size = patch_size
        self.max_patches = max_patches
        self.random_state = random_state

    @_fit_context(prefer_skip_nested_validation=True)
    def fit(self, X, y=None):
        """Only validate the parameters of the estimator.

        This method allows to: (i) validate the parameters of the estimator  and
        (ii) be consistent with the scikit-learn transformer API.

        Parameters
        ----------
        X : ndarray of shape (n_samples, image_height, image_width) or \
                (n_samples, image_height, image_width, n_channels)
            Array of images from which to extract patches. For color images,
            the last dimension specifies the channel: a RGB image would have
            `n_channels=3`.

        y : Ignored
            Not used, present for API consistency by convention.

        Returns
        -------
        self : object
            Returns the instance itself.
        """
        return self

    def transform(self, X):
        """Transform the image samples in `X` into a matrix of patch data.

        Parameters
        ----------
        X : ndarray of shape (n_samples, image_height, image_width) or \
                (n_samples, image_height, image_width, n_channels)
            Array of images from which to extract patches. For color images,
            the last dimension specifies the channel: a RGB image would have
            `n_channels=3`.

        Returns
        -------
        patches : array of shape (n_patches, patch_height, patch_width) or \
                (n_patches, patch_height, patch_width, n_channels)
            The collection of patches extracted from the images, where
            `n_patches` is either `n_samples * max_patches` or the total
            number of patches that can be extracted.
        """
        X = self._validate_data(
            X=X,
            ensure_2d=False,
            allow_nd=True,
            ensure_min_samples=1,
            ensure_min_features=1,
            reset=False,
        )
        random_state = check_random_state(self.random_state)
        n_imgs, img_height, img_width = X.shape[:3]
        if self.patch_size is None:
            patch_size = img_height // 10, img_width // 10
        else:
            if len(self.patch_size) != 2:
                raise ValueError(
                    "patch_size must be a tuple of two integers. Got"
                    f" {self.patch_size} instead."
                )
            patch_size = self.patch_size

        n_imgs, img_height, img_width = X.shape[:3]
        X = np.reshape(X, (n_imgs, img_height, img_width, -1))
        n_channels = X.shape[-1]

        # compute the dimensions of the patches array
        patch_height, patch_width = patch_size
        n_patches = _compute_n_patches(
            img_height, img_width, patch_height, patch_width, self.max_patches
        )
        patches_shape = (n_imgs * n_patches,) + patch_size
        if n_channels > 1:
            patches_shape += (n_channels,)

        # extract the patches
        patches = np.empty(patches_shape)
        for ii, image in enumerate(X):
            patches[ii * n_patches : (ii + 1) * n_patches] = extract_patches_2d(
                image,
                patch_size,
                max_patches=self.max_patches,
                random_state=random_state,
            )
        return patches

    def _more_tags(self):
        return {"X_types": ["3darray"], "stateless": True}