alignment.py

import numpy as np

try:
    import scipy.fft

    my_fftn = scipy.fft.rfftn
    my_ifftn = scipy.fft.irfftn
    my_fft2 = scipy.fft.rfft2
    my_ifft2 = scipy.fft.irfft2

    def my_fft_layout_adapt(x):
        slicelist = tuple(slice(0, (N + 1) // 2) for N in x.shape)
        return x[slicelist]


except ImportError:
    my_fftn = np.fft.fftn
    my_ifftn = np.fft.ifftn
    my_fft2 = np.fft.fft2
    my_ifft2 = np.fft.ifft2

    def my_fft_layout_adapt(x):
        return x


import logging
from numpy.polynomial.polynomial import Polynomial, polyval

from nabu.utils import previouspow2
from nabu.misc import fourier_filters

try:
    from scipy.ndimage.filters import median_filter

    __have_scipy__ = True
except ImportError:
    from silx.math.medianfilter import medfilt2d as median_filter

    __have_scipy__ = False

try:
    import matplotlib.pyplot as plt

    __have_matplotlib__ = True
except ImportError:
    logging.getLogger(__name__).warning("Matplotlib not available. Plotting disabled")

    __have_matplotlib__ = False


class AlignmentBase(object):
    def __init__(self, horz_fft_width=False, verbose=False):
        """
        Alignment basic functions.

        Parameters
        ----------
        horz_fft_width: boolean, optional
            If True, restrict the horizontal size to a power of 2:

            >>> new_x_dim = 2 ** math.floor(math.log2(x_dim))
        verbose: boolean, optional
            When True it will produce verbose output, including plots.
        """
        self._init_parameters(horz_fft_width, verbose)

    def _init_parameters(self, horz_fft_width, verbose):
        self.truncate_horz_pow2 = horz_fft_width

        if verbose and not __have_matplotlib__:
            logging.getLogger(__name__).warning("Matplotlib not available. Plotting disabled, despite being activated by user")
            verbose = False
        self.verbose = verbose

    @staticmethod
    def refine_max_position_2d(f_vals: np.ndarray, fy=None, fx=None):
        """Computes the sub-pixel max position of the given function sampling.

        Parameters
        ----------
        f_vals: numpy.ndarray
            Function values of the sampled points
        fy: numpy.ndarray, optional
            Vertical coordinates of the sampled points
        fx: numpy.ndarray, optional
            Horizontal coordinates of the sampled points

        Raises
        ------
        ValueError
            In case position and values do not have the same size, or in case
            the fitted maximum is outside the fitting region.

        Returns
        -------
        tuple(float, float)
            Estimated (vertical, horizontal) function max, according to the
            coordinates in fy and fx.
        """
        if not (len(f_vals.shape) == 2):
            raise ValueError(
                "The fitted values should form a 2-dimensional array. Array of shape: [%s] was given."
                % (" ".join(("%d" % s for s in f_vals.shape)))
            )
        if fy is None:
            fy_half_size = (f_vals.shape[0] - 1) / 2
            fy = np.linspace(-fy_half_size, fy_half_size, f_vals.shape[0])
        elif not (len(fy.shape) == 1 and np.all(fy.size == f_vals.shape[0])):
            raise ValueError(
                "Vertical coordinates should have the same length as values matrix. Sizes of fy: %d, f_vals: [%s]"
                % (fy.size, " ".join(("%d" % s for s in f_vals.shape)))
            )
        if fx is None:
            fx_half_size = (f_vals.shape[1] - 1) / 2
            fx = np.linspace(-fx_half_size, fx_half_size, f_vals.shape[1])
        elif not (len(fx.shape) == 1 and np.all(fx.size == f_vals.shape[1])):
            raise ValueError(
                "Horizontal coordinates should have the same length as values matrix. Sizes of fx: %d, f_vals: [%s]"
                % (fx.size, " ".join(("%d" % s for s in f_vals.shape)))
            )

        fy, fx = np.meshgrid(fy, fx, indexing="ij")
        fy = fy.flatten()
        fx = fx.flatten()
        coords = np.array([np.ones(f_vals.size), fy, fx, fy * fx, fy ** 2, fx ** 2])

        coeffs = np.linalg.lstsq(coords.T, f_vals.flatten(), rcond=None)[0]

        # For a 1D parabola `f(x) = ax^2 + bx + c`, the vertex position is:
        # x_v = -b / 2a. For a 2D parabola, the vertex position is:
        # (y, x)_v = - b / A, where:
        A = [[2 * coeffs[4], coeffs[3]], [coeffs[3], 2 * coeffs[5]]]
        b = coeffs[1:3]
        vertex_yx = np.linalg.lstsq(A, -b, rcond=None)[0]

        vertex_min_yx = [np.min(fy), np.min(fx)]
        vertex_max_yx = [np.max(fy), np.max(fx)]
        if np.any(vertex_yx < vertex_min_yx) or np.any(vertex_yx > vertex_max_yx):
            raise ValueError(
                "Fitted (y: {}, x: {}) positions are outide the input margins y: [{}, {}], and x: [{}, {}]".format(
                    vertex_yx[0], vertex_yx[1], vertex_min_yx[0], vertex_max_yx[0], vertex_min_yx[1], vertex_max_yx[1],
                )
            )
        return vertex_yx

    @staticmethod
    def refine_max_position_1d(f_vals, fx=None):
        """Computes the sub-pixel max position of the given function sampling.

        Parameters
        ----------
        f_vals: numpy.ndarray
            Function values of the sampled points
        fx: numpy.ndarray, optional
            Coordinates of the sampled points

        Raises
        ------
        ValueError
            In case position and values do not have the same size, or in case
            the fitted maximum is outside the fitting region.

        Returns
        -------
        float
            Estimated function max, according to the coordinates in fx.
        """
        if not len(f_vals.shape) in (1, 2):
            raise ValueError(
                "The fitted values should be either one or a collection of 1-dimensional arrays. Array of shape: [%s] was given."
                % (" ".join(("%d" % s for s in f_vals.shape)))
            )
        num_vals = f_vals.shape[0]

        if fx is None:
            fx_half_size = (num_vals - 1) / 2
            fx = np.linspace(-fx_half_size, fx_half_size, num_vals)
        elif not (len(fx.shape) == 1 and np.all(fx.size == num_vals)):
            raise ValueError(
                "Base coordinates should have the same length as values array. Sizes of fx: %d, f_vals: %d"
                % (fx.size, num_vals)
            )

        if len(f_vals.shape) == 1:
            # using Polynomial.fit, because supposed to be more numerically
            # stable than previous solutions (according to numpy).
            poly = Polynomial.fit(fx, f_vals, deg=2)
            coeffs = poly.convert().coef
        else:
            coords = np.array([np.ones(num_vals), fx, fx ** 2])
            coeffs = np.linalg.lstsq(coords.T, f_vals, rcond=None)[0]

        # For a 1D parabola `f(x) = c + bx + ax^2`, the vertex position is:
        # x_v = -b / 2a.
        vertex_x = -coeffs[1, :] / (2 * coeffs[2, :])

        vertex_min_x = np.min(fx)
        vertex_max_x = np.max(fx)
        lower_bound_ok = vertex_min_x < vertex_x
        upper_bound_ok = vertex_x < vertex_max_x
        if not np.all(lower_bound_ok * upper_bound_ok):
            if len(f_vals.shape) == 1:
                message = "Fitted position {} is outide the input margins [{}, {}]".format(
                    vertex_x, vertex_min_x, vertex_max_x
                )
            else:
                message = "Fitted positions outide the input margins [{}, {}]: %d below and %d above".format(
                    vertex_min_x, vertex_max_x, np.sum(1 - lower_bound_ok), np.sum(1 - upper_bound_ok),
                )
            raise ValueError(message)
        return vertex_x

    @staticmethod
    def extract_peak_region_2d(cc, peak_radius=1, cc_vs=None, cc_hs=None):
        """
        Extracts a region around the maximum value.

        Parameters
        ----------
        cc: numpy.ndarray
            Correlation image.
        peak_radius: int, optional
            The l_inf radius of the area to extract around the peak. The default is 1.
        cc_vs: numpy.ndarray, optional
            The vertical coordinates of `cc`. The default is None.
        cc_hs: numpy.ndarray, optional
            The horizontal coordinates of `cc`. The default is None.

        Returns
        -------
        f_vals: numpy.ndarray
            The extracted function values.
        fv: numpy.ndarray
            The vertical coordinates of the extracted values.
        fh: numpy.ndarray
            The horizontal coordinates of the extracted values.
        """
        img_shape = np.array(cc.shape)
        # get pixel having the maximum value of the correlation array
        pix_max_corr = np.argmax(cc)
        pv, ph = np.unravel_index(pix_max_corr, img_shape)

        # select a n x n neighborhood for the sub-pixel fitting (with wrapping)
        pv = np.arange(pv - peak_radius, pv + peak_radius + 1) % img_shape[-2]
        ph = np.arange(ph - peak_radius, ph + peak_radius + 1) % img_shape[-1]

        # extract the (v, h) pixel coordinates
        fv = None if cc_vs is None else cc_vs[pv]
        fh = None if cc_hs is None else cc_hs[ph]

        # extract the correlation values
        pv, ph = np.meshgrid(pv, ph, indexing="ij")
        f_vals = cc[pv, ph]

        return (f_vals, fv, fh)

    @staticmethod
    def extract_peak_regions_1d(cc, axis=-1, peak_radius=1, cc_coords=None):
        """
        Extracts a region around the maximum value.

        Parameters
        ----------
        cc: numpy.ndarray
            Correlation image.
        axis: int, optional
            Find the max values along the specified direction. The default is -1.
        peak_radius: int, optional
            The l_inf radius of the area to extract around the peak. The default is 1.
        cc_coords: numpy.ndarray, optional
            The coordinates of `cc` along the selected axis. The default is None.

        Returns
        -------
        f_vals: numpy.ndarray
            The extracted function values.
        fc_ax: numpy.ndarray
            The coordinates of the extracted values, along the selected axis.
        """
        img_shape = np.array(cc.shape)
        if not (len(img_shape) == 2):
            raise ValueError(
                "The input image should be a 2-dimensional array. Array of shape: [%s] was given."
                % (" ".join(("%d" % s for s in cc.shape)))
            )
        other_axis = (axis + 1) % 2
        # get pixel having the maximum value of the correlation array
        pix_max = np.argmax(cc, axis=axis)

        # select a n neighborhood for the many 1D sub-pixel fittings (with wrapping)
        p_ax_range = np.arange(-peak_radius, +peak_radius + 1)
        p_ax = (pix_max[None, :] + p_ax_range[:, None]) % img_shape[axis]

        p_ln = np.tile(np.arange(0, img_shape[axis])[None, :], [2 * peak_radius + 1, 1])

        # extract the pixel coordinates along the axis
        fc_ax = None if cc_coords is None else cc_coords[p_ax.flatten()].reshape(p_ax.shape)

        # extract the correlation values
        if other_axis == 0:
            f_vals = cc[p_ln, p_ax]
        else:
            f_vals = cc[p_ax, p_ln]

        return (f_vals, fc_ax)

    def _determine_roi(self, img_shape, roi_yxhw):
        if roi_yxhw is None:
            # vertical window size is reduced to a power of 2 to accelerate fft
            # same thing horizontal window - if requested. Default is not
            roi_yxhw = previouspow2(img_shape)
            if not self.truncate_horz_pow2:
                roi_yxhw[1] = img_shape[1]

        if len(roi_yxhw) == 2:  # Convert centered 2-element roi into 4-element
            roi_yxhw = np.array(roi_yxhw, dtype=np.int)
            roi_yxhw = np.concatenate(((img_shape - roi_yxhw) // 2, roi_yxhw))
        return roi_yxhw

    @staticmethod
    def _prepare_image(
        img, invalid_val=1e-5, roi_yxhw=None, median_filt_shape=None, low_pass=None, high_pass=None,
    ):
        """
        Prepare and returns  a cropped  and filtered image, or array of filtered images if the input is an  array of images.

        Parameters
        ----------
        img: numpy.ndarray
            image or stack of images
        invalid_val: float
            value to be used in replacement of nan and inf values
        median_filt_shape: int or sequence of int
            the width or the widths of the median window
        low_pass: float or sequence of two floats
            Low-pass filter properties, as described in `nabu.misc.fourier_filters`
        high_pass: float or sequence of two floats
            High-pass filter properties, as described in `nabu.misc.fourier_filters`

        Returns
        -------
        numpy.array_like
            The computed filter
        """
        img = np.squeeze(img)  # Removes singleton dimensions, but does a shallow copy
        img = np.ascontiguousarray(img)

        if roi_yxhw is not None:
            img = img[
                ..., roi_yxhw[0] : roi_yxhw[0] + roi_yxhw[2], roi_yxhw[1] : roi_yxhw[1] + roi_yxhw[3],
            ]

        img = img.copy()

        img[np.isnan(img)] = invalid_val
        img[np.isinf(img)] = invalid_val

        if high_pass is not None or low_pass is not None:
            img_filter = fourier_filters.get_bandpass_filter(
                img.shape[-2:], cutoff_lowpass=low_pass, cutoff_highpass=high_pass
            )
            # fft2 and iff2 use axes=(-2, -1) by default
            img = my_ifft2(my_fft2(img) * my_fft_layout_adapt(img_filter)).real

        if median_filt_shape is not None:
            img_shape = img.shape
            if __have_scipy__:
                # expanding filter shape with ones, to cover the stack of images
                # but disabling inter-image filtering
                median_filt_shape = np.concatenate(
                    (np.ones((len(img_shape) - len(median_filt_shape),), dtype=np.int), median_filt_shape,)
                )
                img = median_filter(img, size=median_filt_shape)
            else:
                if len(img_shape) == 2:
                    img = median_filter(img, kernel_size=median_filt_shape)
                elif len(img_shape) > 2:
                    # if dealing with a stack of images, we have to do them one by one
                    img = np.reshape(img, (-1,) + tuple(img_shape[-2:]))
                    for ii in range(img.shape[0]):
                        img[ii, ...] = median_filter(img[ii, ...], kernel_size=median_filt_shape)
                    img = np.reshape(img, img_shape)

        return img

    @staticmethod
    def _compute_correlation_fft(img_1, img_2, padding_mode, axes=(-2, -1), low_pass=None, high_pass=None):
        do_circular_conv = padding_mode is None or padding_mode == "wrap"
        if not do_circular_conv:
            img_shape = img_2.shape
            pad_size = np.ceil(np.array(img_shape) / 2).astype(np.int)
            pad_array = [(0,)] * len(img_shape)
            for a in axes:
                pad_array[a] = (pad_size[a],)

            img_1 = np.pad(img_1, pad_array, mode=padding_mode)
            img_2 = np.pad(img_2, pad_array, mode=padding_mode)

        # compute fft's of the 2 images
        img_fft_1 = my_fftn(img_1, axes=axes)
        img_fft_2 = np.conjugate(my_fftn(img_2, axes=axes))

        img_prod = img_fft_1 * img_fft_2

        if low_pass is not None or high_pass is not None:
            filt = fourier_filters.get_bandpass_filter(img_prod.shape[-2:], cutoff_lowpass=low_pass, cutoff_highpass=high_pass)
            img_prod *= filt

        # inverse fft of the product to get cross_correlation of the 2 images
        cc = np.real(my_ifftn(img_prod, axes=axes))

        if not do_circular_conv:
            cc = np.fft.fftshift(cc, axes=axes)

            slicing = [slice(None)] * len(img_shape)
            for a in axes:
                slicing[a] = slice(pad_size[a], cc.shape[a] - pad_size[a])
            cc = cc[tuple(slicing)]

            cc = np.fft.ifftshift(cc, axes=axes)

        return cc


class CenterOfRotation(AlignmentBase):
    @staticmethod
    def _check_img_sizes(img_1: np.ndarray, img_2: np.ndarray):
        shape_1 = np.squeeze(img_1).shape
        shape_2 = np.squeeze(img_2).shape
        if not len(shape_1) == 2:
            raise ValueError(
                "Images need to be 2-dimensional. Shape of image #1: %s" % (" ".join(("%d" % x for x in shape_1)))
            )
        if not len(shape_2) == 2:
            raise ValueError(
                "Images need to be 2-dimensional. Shape of image #2: %s" % (" ".join(("%d" % x for x in shape_2)))
            )
        if not np.all(shape_1 == shape_2):
            raise ValueError(
                "Images need to be of the same shape. Shape of image #1: %s, image #2: %s"
                % (" ".join(("%d" % x for x in shape_1)), " ".join(("%d" % x for x in shape_2)),)
            )

    def find_shift(
        self,
        img_1: np.ndarray,
        img_2: np.ndarray,
        roi_yxhw=None,
        median_filt_shape=None,
        padding_mode=None,
        peak_fit_radius=1,
        high_pass=None,
        low_pass=None,
    ):
        """Find the Center of Rotation (CoR), given to images.

        This method finds the half-shift between two opposite images, by
        means of correlation computed in Fourier space.

        The output of this function, allows to compute motor movements for
        aligning the sample rotation axis. Given the following values:

        - L1: distance from source to motor
        - L2: distance from source to detector
        - ps: physical pixel size
        - v: output of this function

        displacement of motor = (L1 / L2 * ps) * v

        Parameters
        ----------
        img_1: numpy.ndarray
            First image
        img_2: numpy.ndarray
            Second image, it needs to have been flipped already (e.g. using numpy.fliplr).
        roi_yxhw: (2, ) or (4, ) numpy.ndarray, tuple, or array, optional
            4 elements vector containing: vertical and horizontal coordinates
            of first pixel, plus height and width of the Region of Interest (RoI).
            Or a 2 elements vector containing: plus height and width of the
            centered Region of Interest (RoI).
            Default is None -> deactivated.
        median_filt_shape: (2, ) numpy.ndarray, tuple, or array, optional
            Shape of the median filter window. Default is None -> deactivated.
        padding_mode: str in numpy.pad's mode list, optional
            Padding mode, which determines the type of convolution. If None or
            'wrap' are passed, this resorts to the traditional circular convolution.
            If 'edge' or 'constant' are passed, it results in a linear convolution.
            Default is the circular convolution.
            All options are:
                None | 'constant' | 'edge' | 'linear_ramp' | 'maximum' | 'mean'
                | 'median' | 'minimum' | 'reflect' | 'symmetric' |'wrap'
        peak_fit_radius: int, optional
            Radius size around the max correlation pixel, for sub-pixel fitting.
            Minimum and default value is 1.
        low_pass: float or sequence of two floats
            Low-pass filter properties, as described in `nabu.misc.fourier_filters`
        high_pass: float or sequence of two floats
            High-pass filter properties, as described in `nabu.misc.fourier_filters`

        Raises
        ------
        ValueError
            In case images are not 2-dimensional or have different sizes.

        Returns
        -------
        float
            Estimated center of rotation position from the center of the RoI in pixels.

        Examples
        --------
        The following code computes the center of rotation position for two
        given images in a tomography scan, where the second image is taken at
        180 degrees from the first.

        >>> radio1 = data[0, :, :]
        ... radio2 = np.fliplr(data[1, :, :])
        ... CoR_calc = CenterOfRotation()
        ... cor_position = CoR_calc.find_shift(radio1, radio2)

        Or for noisy images:

        >>> cor_position = CoR_calc.find_shift(radio1, radio2, median_filt_shape=(3, 3))
        """
        self._check_img_sizes(img_1, img_2)

        if peak_fit_radius < 1:
            logging.getLogger(__name__).warning(
                "Parameter peak_fit_radius should be at least 1, given: %d instead." % peak_fit_radius
            )
            peak_fit_radius = 1

        img_shape = img_2.shape
        roi_yxhw = self._determine_roi(img_shape, roi_yxhw)

        img_1 = self._prepare_image(img_1, roi_yxhw=roi_yxhw, median_filt_shape=median_filt_shape)
        img_2 = self._prepare_image(img_2, roi_yxhw=roi_yxhw, median_filt_shape=median_filt_shape)

        cc = self._compute_correlation_fft(img_1, img_2, padding_mode, high_pass=high_pass, low_pass=low_pass)

        img_shape = img_2.shape
        cc_vs = np.fft.fftfreq(img_shape[-2], 1 / img_shape[-2])
        cc_hs = np.fft.fftfreq(img_shape[-1], 1 / img_shape[-1])

        (f_vals, fv, fh) = self.extract_peak_region_2d(cc, peak_radius=peak_fit_radius, cc_vs=cc_vs, cc_hs=cc_hs)
        fitted_shifts_vh = self.refine_max_position_2d(f_vals, fv, fh)

        return fitted_shifts_vh[-1] / 2.0

    __call__ = find_shift


class DetectorTranslationAlongBeam(AlignmentBase):
    @staticmethod
    def _check_img_sizes(img_stack: np.ndarray, img_pos: np.ndarray):
        shape_stack = np.squeeze(img_stack).shape
        shape_pos = np.squeeze(img_pos).shape
        if not len(shape_stack) == 3:
            raise ValueError(
                "A stack of 2-dimensional images is required. Shape of stack: %s" % (" ".join(("%d" % x for x in shape_stack)))
            )
        if not len(shape_pos) == 1:
            raise ValueError(
                "Positions need to be a 1-dimensional array. Shape of the positions variable: %s"
                % (" ".join(("%d" % x for x in shape_pos)))
            )
        if not shape_stack[0] == shape_pos[0]:
            raise ValueError(
                "The same number of images and positions is required."
                + " Shape of stack: %s, shape of positions variable: %s"
                % (" ".join(("%d" % x for x in shape_stack)), " ".join(("%d" % x for x in shape_pos)),)
            )

    def find_shift(
        self,
        img_stack: np.ndarray,
        img_pos: np.array,
        roi_yxhw=None,
        median_filt_shape=None,
        padding_mode=None,
        peak_fit_radius=1,
        high_pass=None,
        low_pass=None,
        equispaced_increments=True,
        return_shifts=False,
        use_adjacent_imgs=False,
    ):
        """Find  vertical and  horizontal position increments per a unit-distance detector translation along the
        traslation axis. The units are pixel_unit/input_unit where input_unit are the  unit that the user has  used
        to pass the argument img_pos. The output expresses shifts of the detector so that if the image is moving
        in the positive direction (expressed in pixels coordinates) the output will be negative because it means
        that the detector as a whole is shifting in the opposite direction (taking the shaped  beam as a reference)

        Parameters
        ----------
        img_stack: numpy.ndarray
            A stack of images (usually 4) at different distances
        img_pos: numpy.ndarray
            Position of the images along the translation axis
        roi_yxhw: (2, ) or (4, ) numpy.ndarray, tuple, or array, optional
            4 elements vector containing: vertical and horizontal coordinates
            of first pixel, plus height and width of the Region of Interest (RoI).
            Or a 2 elements vector containing: plus height and width of the
            centered Region of Interest (RoI).
            Default is None -> deactivated.
        median_filt_shape: (2, ) numpy.ndarray, tuple, or array, optional
            Shape of the median filter window. Default is None -> deactivated.
        padding_mode: str in numpy.pad's mode list, optional
            Padding mode, which determines the type of convolution. If None or
            'wrap' are passed, this resorts to the traditional circular convolution.
            If 'edge' or 'constant' are passed, it results in a linear convolution.
            Default is the circular convolution.
            All options are:
                None | 'constant' | 'edge' | 'linear_ramp' | 'maximum' | 'mean'
                | 'median' | 'minimum' | 'reflect' | 'symmetric' |'wrap'
        peak_fit_radius: int, optional
            Radius size around the max correlation pixel, for sub-pixel fitting.
            Minimum and default value is 1.
        low_pass: float or sequence of two floats
            Low-pass filter properties, as described in `nabu.misc.fourier_filters`.
        high_pass: float or sequence of two floats
            High-pass filter properties, as described in `nabu.misc.fourier_filters`.
        return_shifts: boolean, optional
            if a True a list, containing for each given image of the stack a tuple of shifts, will
            be return together with  the increment per unit-distance.
            This slot is intended for introspection.
        use_adjacent_imgs: boolean, optional
            Compute correlation between adjacent images. It is better when dealing with large shifts.

        Returns
        -------
        tuple(float, float)
            Estimated (vertical, horizontal) increment per unit-distance of the
            ratio between pixel-size and detector translation.
            Optionally a tuple containing the shifts per each image are found if optional argument
            return_shifts is true

        Examples
        --------

        >>> import numpy as np
        ... import scipy.ndimage
        ... from nabu.preproc.alignment import  DetectorTranslationAlongBeam
        ...
        ... T_calc = DetectorTranslationAlongBeam()
        ...
        ... stack = np.zeros([4, 512, 512], "d")
        ...
        ... # Add low frequency spurious component
        ... for i in range(4):
        ...     stack[i, 200 - i * 10, 200 - i * 10] = 1
        ... stack = scipy.ndimage.filters.gaussian_filter(stack, [0, 10, 10.0]) * 100
        ...
        ... # Add the feature
        ... x, y = np.meshgrid(np.arange(stack.shape[-1]), np.arange(stack.shape[-2]))
        ... for i in range(4):
        ...     xc = x - (250 + i * 1.234)
        ...     yc = y - (250 + i * 1.234 * 2)
        ...     stack[i] += np.exp(-(xc * xc + yc * yc) * 0.5)
        ...
        ... # Find the shifts from the features
        ... shifts_v, shifts_h, found_shifts_list = T_calc.find_shift(
        ...     stack, np.array([0.0, 1, 2, 3]), high_pass=1.0, return_shifts=True
        ... )
        ... print(shifts_v, shifts_h)
        >>> ( -2.47 , -1.236 )



        """
        self._check_img_sizes(img_stack, img_pos)

        if peak_fit_radius < 1:
            logging.getLogger(__name__).warning(
                "Parameter peak_fit_radius should be at least 1, given: %d instead." % peak_fit_radius
            )
            peak_fit_radius = 1

        num_imgs = img_stack.shape[0]
        img_shape = img_stack.shape[-2:]
        roi_yxhw = self._determine_roi(img_shape, roi_yxhw)

        img_stack = self._prepare_image(img_stack, roi_yxhw=roi_yxhw, median_filt_shape=median_filt_shape)

        # do correlations
        ccs = [
            self._compute_correlation_fft(
                img_stack[ii - 1 if use_adjacent_imgs else 0, ...],
                img_stack[ii, ...],
                padding_mode,
                high_pass=high_pass,
                low_pass=low_pass,
            )
            for ii in range(1, num_imgs)
        ]

        img_shape = img_stack.shape[-2:]
        cc_vs = np.fft.fftfreq(img_shape[-2], 1 / img_shape[-2])
        cc_hs = np.fft.fftfreq(img_shape[-1], 1 / img_shape[-1])

        shifts_vh = np.empty((num_imgs - 1, 2))
        for ii, cc in enumerate(ccs):
            (f_vals, fv, fh) = self.extract_peak_region_2d(cc, peak_radius=peak_fit_radius, cc_vs=cc_vs, cc_hs=cc_hs)
            shifts_vh[ii, :] = self.refine_max_position_2d(f_vals, fv, fh)

        if use_adjacent_imgs:
            shifts_vh = np.cumsum(shifts_vh, axis=0)

        img_shifts_vh = np.concatenate(([[0, 0]], shifts_vh), axis=0)

        # Polynomial.fit is supposed to be more numerically stable than polyfit
        # (according to numpy)
        coeffs_v = Polynomial.fit(img_pos, img_shifts_vh[:, 0], deg=1).convert().coef
        coeffs_h = Polynomial.fit(img_pos, img_shifts_vh[:, 1], deg=1).convert().coef

        if self.verbose:
            f, axs = plt.subplots(1, 2)
            axs[0].scatter(img_pos, img_shifts_vh[:, 0])
            axs[0].plot(img_pos, polyval(img_pos, coeffs_v))
            axs[0].set_title("Vertical shifts")
            axs[1].scatter(img_pos, img_shifts_vh[:, 1])
            axs[1].plot(img_pos, polyval(img_pos, coeffs_h))
            axs[1].set_title("Horizontal shifts")
            plt.show(block=False)

        if return_shifts:
            return coeffs_v[1], coeffs_h[1], shifts_vh
        else:
            return coeffs_v[1], coeffs_h[1]