Source code for dipy.align.metrics

"""Metrics for Symmetric Diffeomorphic Registration"""

import abc

import numpy as np
from numpy import gradient
from scipy import ndimage

from dipy.align import (
    crosscorr as cc,
    expectmax as em,
    floating,
    sumsqdiff as ssd,
    vector_fields as vfu,
)
from dipy.testing.decorators import warning_for_keywords


[docs] class SimilarityMetric: def __init__(self, dim): r"""Similarity Metric abstract class A similarity metric is in charge of keeping track of the numerical value of the similarity (or distance) between the two given images. It also computes the update field for the forward and inverse displacement fields to be used in a gradient-based optimization algorithm. Note that this metric does not depend on any transformation (affine or non-linear) so it assumes the static and moving images are already warped Parameters ---------- dim : int (either 2 or 3) the dimension of the image domain """ self.dim = dim self.levels_above = None self.levels_below = None self.static_image = None self.static_affine = None self.static_spacing = None self.static_direction = None self.moving_image = None self.moving_affine = None self.moving_spacing = None self.moving_direction = None self.mask0 = False
[docs] def set_levels_below(self, levels): r"""Informs the metric how many pyramid levels are below the current one Informs this metric the number of pyramid levels below the current one. The metric may change its behavior (e.g. number of inner iterations) accordingly Parameters ---------- levels : int the number of levels below the current Gaussian Pyramid level """ self.levels_below = levels
[docs] def set_levels_above(self, levels): r"""Informs the metric how many pyramid levels are above the current one Informs this metric the number of pyramid levels above the current one. The metric may change its behavior (e.g. number of inner iterations) accordingly Parameters ---------- levels : int the number of levels above the current Gaussian Pyramid level """ self.levels_above = levels
[docs] def set_static_image( self, static_image, static_affine, static_spacing, static_direction ): r"""Sets the static image being compared against the moving one. Sets the static image. The default behavior (of this abstract class) is simply to assign the reference to an attribute, but generalizations of the metric may need to perform other operations Parameters ---------- static_image : array, shape (R, C) or (S, R, C) the static image """ self.static_image = static_image self.static_affine = static_affine self.static_spacing = static_spacing self.static_direction = static_direction
[docs] def use_static_image_dynamics(self, original_static_image, transformation): r"""This is called by the optimizer just after setting the static image. This method allows the metric to compute any useful information from knowing how the current static image was generated (as the transformation of an original static image). This method is called by the optimizer just after it sets the static image. Transformation will be an instance of DiffeomorficMap or None if the original_static_image equals self.moving_image. Parameters ---------- original_static_image : array, shape (R, C) or (S, R, C) original image from which the current static image was generated transformation : DiffeomorphicMap object the transformation that was applied to original image to generate the current static image """ pass
[docs] def set_moving_image( self, moving_image, moving_affine, moving_spacing, moving_direction ): r"""Sets the moving image being compared against the static one. Sets the moving image. The default behavior (of this abstract class) is simply to assign the reference to an attribute, but generalizations of the metric may need to perform other operations Parameters ---------- moving_image : array, shape (R, C) or (S, R, C) the moving image """ self.moving_image = moving_image self.moving_affine = moving_affine self.moving_spacing = moving_spacing self.moving_direction = moving_direction
[docs] def use_moving_image_dynamics(self, original_moving_image, transformation): r"""This is called by the optimizer just after setting the moving image This method allows the metric to compute any useful information from knowing how the current static image was generated (as the transformation of an original static image). This method is called by the optimizer just after it sets the static image. Transformation will be an instance of DiffeomorficMap or None if the original_moving_image equals self.moving_image. Parameters ---------- original_moving_image : array, shape (R, C) or (S, R, C) original image from which the current moving image was generated transformation : DiffeomorphicMap object the transformation that was applied to the original image to generate the current moving image """ pass
[docs] @abc.abstractmethod def initialize_iteration(self): r"""Prepares the metric to compute one displacement field iteration. This method will be called before any compute_forward or compute_backward call, this allows the Metric to pre-compute any useful information for speeding up the update computations. This initialization was needed in ANTS because the updates are called once per voxel. In Python this is unpractical, though. """
[docs] @abc.abstractmethod def free_iteration(self): r"""Releases the resources no longer needed by the metric This method is called by the RegistrationOptimizer after the required iterations have been computed (forward and / or backward) so that the SimilarityMetric can safely delete any data it computed as part of the initialization """
[docs] @abc.abstractmethod def compute_forward(self): r"""Computes one step bringing the reference image towards the static. Computes the forward update field to register the moving image towards the static image in a gradient-based optimization algorithm """
[docs] @abc.abstractmethod def compute_backward(self): r"""Computes one step bringing the static image towards the moving. Computes the backward update field to register the static image towards the moving image in a gradient-based optimization algorithm """
[docs] @abc.abstractmethod def get_energy(self): r"""Numerical value assigned by this metric to the current image pair Must return the numeric value of the similarity between the given static and moving images """
[docs] class CCMetric(SimilarityMetric): @warning_for_keywords() def __init__(self, dim, *, sigma_diff=2.0, radius=4): r"""Normalized Cross-Correlation Similarity metric. Parameters ---------- dim : int (either 2 or 3) the dimension of the image domain sigma_diff : the standard deviation of the Gaussian smoothing kernel to be applied to the update field at each iteration radius : int the radius of the squared (cubic) neighborhood at each voxel to be considered to compute the cross correlation """ super(CCMetric, self).__init__(dim) self.sigma_diff = sigma_diff self.radius = radius self._connect_functions() def _connect_functions(self): r"""Assign the methods to be called according to the image dimension Assigns the appropriate functions to be called for precomputing the cross-correlation factors according to the dimension of the input images """ if self.dim == 2: self.precompute_factors = cc.precompute_cc_factors_2d self.compute_forward_step = cc.compute_cc_forward_step_2d self.compute_backward_step = cc.compute_cc_backward_step_2d self.reorient_vector_field = vfu.reorient_vector_field_2d elif self.dim == 3: self.precompute_factors = cc.precompute_cc_factors_3d self.compute_forward_step = cc.compute_cc_forward_step_3d self.compute_backward_step = cc.compute_cc_backward_step_3d self.reorient_vector_field = vfu.reorient_vector_field_3d else: raise ValueError(f"CC Metric not defined for dim. {self.dim}")
[docs] def initialize_iteration(self): r"""Prepares the metric to compute one displacement field iteration. Pre-computes the cross-correlation factors for efficient computation of the gradient of the Cross Correlation w.r.t. the displacement field. It also pre-computes the image gradients in the physical space by re-orienting the gradients in the voxel space using the corresponding affine transformations. """ min_size = self.radius * 2 + 1 def invalid_image_size(image): return any(size < min_size for size in image.shape) msg = ( "Each image dimension should be superior to 2 * radius + 1 " f"({min_size}). Decrease CCMetric radius ({self.radius}) or " "increase your image size (shape=%(shape)s)." ) if invalid_image_size(self.static_image): raise ValueError( "Static image size is too small. " + msg % {"shape": self.static_image.shape} ) if invalid_image_size(self.moving_image): raise ValueError( "Moving image size is too small. " + msg % {"shape": self.moving_image.shape} ) self.factors = self.precompute_factors( self.static_image, self.moving_image, self.radius ) self.factors = np.array(self.factors) self.gradient_moving = np.empty( shape=self.moving_image.shape + (self.dim,), dtype=floating ) for i, grad in enumerate(gradient(self.moving_image)): self.gradient_moving[..., i] = grad # Convert moving image's gradient field from voxel to physical space if self.moving_spacing is not None: self.gradient_moving /= self.moving_spacing if self.moving_direction is not None: self.reorient_vector_field(self.gradient_moving, self.moving_direction) self.gradient_static = np.empty( shape=self.static_image.shape + (self.dim,), dtype=floating ) for i, grad in enumerate(gradient(self.static_image)): self.gradient_static[..., i] = grad # Convert moving image's gradient field from voxel to physical space if self.static_spacing is not None: self.gradient_static /= self.static_spacing if self.static_direction is not None: self.reorient_vector_field(self.gradient_static, self.static_direction)
[docs] def free_iteration(self): r"""Frees the resources allocated during initialization""" del self.factors del self.gradient_moving del self.gradient_static
[docs] def compute_forward(self): r"""Computes one step bringing the moving image towards the static. Computes the update displacement field to be used for registration of the moving image towards the static image """ displacement, self.energy = self.compute_forward_step( self.gradient_static, self.factors, self.radius ) displacement = np.array(displacement) for i in range(self.dim): displacement[..., i] = ndimage.gaussian_filter( displacement[..., i], self.sigma_diff ) return displacement
[docs] def compute_backward(self): r"""Computes one step bringing the static image towards the moving. Computes the update displacement field to be used for registration of the static image towards the moving image """ displacement, energy = self.compute_backward_step( self.gradient_moving, self.factors, self.radius ) displacement = np.array(displacement) for i in range(self.dim): displacement[..., i] = ndimage.gaussian_filter( displacement[..., i], self.sigma_diff ) return displacement
[docs] def get_energy(self): r"""Numerical value assigned by this metric to the current image pair Returns the Cross Correlation (data term) energy computed at the largest iteration """ return self.energy
[docs] class EMMetric(SimilarityMetric): @warning_for_keywords() def __init__( self, dim, *, smooth=1.0, inner_iter=5, q_levels=256, double_gradient=True, step_type="gauss_newton", ): r"""Expectation-Maximization Metric Similarity metric based on the Expectation-Maximization algorithm to handle multi-modal images. The transfer function is modeled as a set of hidden random variables that are estimated at each iteration of the algorithm. Parameters ---------- dim : int (either 2 or 3) the dimension of the image domain smooth : float smoothness parameter, the larger the value the smoother the deformation field inner_iter : int number of iterations to be performed at each level of the multi- resolution Gauss-Seidel optimization algorithm (this is not the number of steps per Gaussian Pyramid level, that parameter must be set for the optimizer, not the metric) q_levels : number of quantization levels (equal to the number of hidden variables in the EM algorithm) double_gradient : boolean if True, the gradient of the expected static image under the moving modality will be added to the gradient of the moving image, similarly, the gradient of the expected moving image under the static modality will be added to the gradient of the static image. step_type : string ('gauss_newton', 'demons') the optimization schedule to be used in the multi-resolution Gauss-Seidel optimization algorithm (not used if Demons Step is selected) """ super(EMMetric, self).__init__(dim) self.smooth = smooth self.inner_iter = inner_iter self.q_levels = q_levels self.use_double_gradient = double_gradient self.step_type = step_type self.static_image_mask = None self.moving_image_mask = None self.staticq_means_field = None self.movingq_means_field = None self.movingq_levels = None self.staticq_levels = None self._connect_functions() def _connect_functions(self): r"""Assign the methods to be called according to the image dimension Assigns the appropriate functions to be called for image quantization, statistics computation and multi-resolution iterations according to the dimension of the input images """ if self.dim == 2: self.quantize = em.quantize_positive_2d self.compute_stats = em.compute_masked_class_stats_2d self.reorient_vector_field = vfu.reorient_vector_field_2d elif self.dim == 3: self.quantize = em.quantize_positive_3d self.compute_stats = em.compute_masked_class_stats_3d self.reorient_vector_field = vfu.reorient_vector_field_3d else: raise ValueError(f"EM Metric not defined for dim. {self.dim}") if self.step_type == "demons": self.compute_step = self.compute_demons_step elif self.step_type == "gauss_newton": self.compute_step = self.compute_gauss_newton_step else: raise ValueError(f"Opt. step {self.step_type} not defined")
[docs] def initialize_iteration(self): r"""Prepares the metric to compute one displacement field iteration. Pre-computes the transfer functions (hidden random variables) and variances of the estimators. Also pre-computes the gradient of both input images. Note that once the images are transformed to the opposite modality, the gradient of the transformed images can be used with the gradient of the corresponding modality in the same fashion as diff-demons does for mono-modality images. If the flag self.use_double_gradient is True these gradients are averaged. """ sampling_mask = self.static_image_mask * self.moving_image_mask self.sampling_mask = sampling_mask staticq, self.staticq_levels, hist = self.quantize( self.static_image, self.q_levels ) staticq = np.array(staticq, dtype=np.int32) self.staticq_levels = np.array(self.staticq_levels) staticq_means, staticq_vars = self.compute_stats( sampling_mask, self.moving_image, self.q_levels, staticq ) staticq_means[0] = 0 self.staticq_means = np.array(staticq_means) self.staticq_variances = np.array(staticq_vars) self.staticq_sigma_sq_field = self.staticq_variances[staticq] self.staticq_means_field = self.staticq_means[staticq] self.gradient_moving = np.empty( shape=self.moving_image.shape + (self.dim,), dtype=floating ) for i, grad in enumerate(gradient(self.moving_image)): self.gradient_moving[..., i] = grad # Convert moving image's gradient field from voxel to physical space if self.moving_spacing is not None: self.gradient_moving /= self.moving_spacing if self.moving_direction is not None: self.reorient_vector_field(self.gradient_moving, self.moving_direction) self.gradient_static = np.empty( shape=self.static_image.shape + (self.dim,), dtype=floating ) for i, grad in enumerate(gradient(self.static_image)): self.gradient_static[..., i] = grad # Convert moving image's gradient field from voxel to physical space if self.static_spacing is not None: self.gradient_static /= self.static_spacing if self.static_direction is not None: self.reorient_vector_field(self.gradient_static, self.static_direction) movingq, self.movingq_levels, hist = self.quantize( self.moving_image, self.q_levels ) movingq = np.array(movingq, dtype=np.int32) self.movingq_levels = np.array(self.movingq_levels) movingq_means, movingq_variances = self.compute_stats( sampling_mask, self.static_image, self.q_levels, movingq ) movingq_means[0] = 0 self.movingq_means = np.array(movingq_means) self.movingq_variances = np.array(movingq_variances) self.movingq_sigma_sq_field = self.movingq_variances[movingq] self.movingq_means_field = self.movingq_means[movingq] if self.use_double_gradient: for i, grad in enumerate(gradient(self.staticq_means_field)): self.gradient_moving[..., i] += grad for i, grad in enumerate(gradient(self.movingq_means_field)): self.gradient_static[..., i] += grad
[docs] def free_iteration(self): r""" Frees the resources allocated during initialization """ del self.sampling_mask del self.staticq_levels del self.movingq_levels del self.staticq_sigma_sq_field del self.staticq_means_field del self.movingq_sigma_sq_field del self.movingq_means_field del self.gradient_moving del self.gradient_static
[docs] def compute_forward(self): """Computes one step bringing the reference image towards the static. Computes the forward update field to register the moving image towards the static image in a gradient-based optimization algorithm """ return self.compute_step(forward_step=True)
[docs] def compute_backward(self): r"""Computes one step bringing the static image towards the moving. Computes the update displacement field to be used for registration of the static image towards the moving image """ return self.compute_step(forward_step=False)
[docs] @warning_for_keywords() def compute_gauss_newton_step(self, *, forward_step=True): r"""Computes the Gauss-Newton energy minimization step Computes the Newton step to minimize this energy, i.e., minimizes the linearized energy function with respect to the regularized displacement field (this step does not require post-smoothing, as opposed to the demons step, which does not include regularization). To accelerate convergence we use the multi-grid Gauss-Seidel algorithm proposed by :footcite:t:`Bruhn2005`. Parameters ---------- forward_step : boolean if True, computes the Newton step in the forward direction (warping the moving towards the static image). If False, computes the backward step (warping the static image to the moving image) Returns ------- displacement : array, shape (R, C, 2) or (S, R, C, 3) the Newton step References ---------- .. footbibliography:: """ reference_shape = self.static_image.shape if forward_step: gradient = self.gradient_static delta = self.staticq_means_field - self.moving_image sigma_sq_field = self.staticq_sigma_sq_field else: gradient = self.gradient_moving delta = self.movingq_means_field - self.static_image sigma_sq_field = self.movingq_sigma_sq_field displacement = np.zeros(shape=reference_shape + (self.dim,), dtype=floating) if self.dim == 2: self.energy = v_cycle_2d( self.levels_below, self.inner_iter, delta, sigma_sq_field, gradient, None, self.smooth, displacement, ) else: self.energy = v_cycle_3d( self.levels_below, self.inner_iter, delta, sigma_sq_field, gradient, None, self.smooth, displacement, ) return displacement
[docs] @warning_for_keywords() def compute_demons_step(self, *, forward_step=True): r"""Demons step for EM metric Parameters ---------- forward_step : boolean if True, computes the Demons step in the forward direction (warping the moving towards the static image). If False, computes the backward step (warping the static image to the moving image) Returns ------- displacement : array, shape (R, C, 2) or (S, R, C, 3) the Demons step """ sigma_reg_2 = np.sum(self.static_spacing**2) / self.dim if forward_step: gradient = self.gradient_static delta_field = self.static_image - self.movingq_means_field sigma_sq_field = self.movingq_sigma_sq_field else: gradient = self.gradient_moving delta_field = self.moving_image - self.staticq_means_field sigma_sq_field = self.staticq_sigma_sq_field if self.dim == 2: step, self.energy = em.compute_em_demons_step_2d( delta_field, sigma_sq_field, gradient, sigma_reg_2, None ) else: step, self.energy = em.compute_em_demons_step_3d( delta_field, sigma_sq_field, gradient, sigma_reg_2, None ) for i in range(self.dim): step[..., i] = ndimage.gaussian_filter(step[..., i], self.smooth) return step
[docs] def get_energy(self): r"""The numerical value assigned by this metric to the current image pair Returns the EM (data term) energy computed at the largest iteration """ return self.energy
[docs] def use_static_image_dynamics(self, original_static_image, transformation): r"""This is called by the optimizer just after setting the static image. EMMetric takes advantage of the image dynamics by computing the current static image mask from the originalstaticImage mask (warped by nearest neighbor interpolation) Parameters ---------- original_static_image : array, shape (R, C) or (S, R, C) the original static image from which the current static image was generated, the current static image is the one that was provided via 'set_static_image(...)', which may not be the same as the original static image but a warped version of it (even the static image changes during Symmetric Normalization, not only the moving one). transformation : DiffeomorphicMap object the transformation that was applied to the original_static_image to generate the current static image """ self.static_image_mask = (original_static_image > 0).astype(np.int32) if transformation is None: return shape = np.array(self.static_image.shape, dtype=np.int32) affine = self.static_affine self.static_image_mask = transformation.transform( self.static_image_mask, interpolation="nearest", image_world2grid=None, out_shape=shape, out_grid2world=affine, )
[docs] def use_moving_image_dynamics(self, original_moving_image, transformation): r"""This is called by the optimizer just after setting the moving image. EMMetric takes advantage of the image dynamics by computing the current moving image mask from the original_moving_image mask (warped by nearest neighbor interpolation) Parameters ---------- original_moving_image : array, shape (R, C) or (S, R, C) the original moving image from which the current moving image was generated, the current moving image is the one that was provided via 'set_moving_image(...)', which may not be the same as the original moving image but a warped version of it. transformation : DiffeomorphicMap object the transformation that was applied to the original_moving_image to generate the current moving image """ self.moving_image_mask = (original_moving_image > 0).astype(np.int32) if transformation is None: return shape = np.array(self.moving_image.shape, dtype=np.int32) affine = self.moving_affine self.moving_image_mask = transformation.transform( self.moving_image_mask, interpolation="nearest", image_world2grid=None, out_shape=shape, out_grid2world=affine, )
[docs] class SSDMetric(SimilarityMetric): @warning_for_keywords() def __init__(self, dim, *, smooth=4, inner_iter=10, step_type="demons"): r"""Sum of Squared Differences (SSD) Metric Similarity metric for (mono-modal) nonlinear image registration defined by the sum of squared differences (SSD) Parameters ---------- dim : int (either 2 or 3) the dimension of the image domain smooth : float smoothness parameter, the larger the value the smoother the deformation field inner_iter : int number of iterations to be performed at each level of the multi- resolution Gauss-Seidel optimization algorithm (this is not the number of steps per Gaussian Pyramid level, that parameter must be set for the optimizer, not the metric) step_type : string the displacement field step to be computed when 'compute_forward' and 'compute_backward' are called. Either 'demons' or 'gauss_newton' """ super(SSDMetric, self).__init__(dim) self.smooth = smooth self.inner_iter = inner_iter self.step_type = step_type self.levels_below = 0 self._connect_functions() def _connect_functions(self): r"""Assign the methods to be called according to the image dimension Assigns the appropriate functions to be called for vector field reorientation and displacement field steps according to the dimension of the input images and the select type of step (either Demons or Gauss Newton) """ if self.dim == 2: self.reorient_vector_field = vfu.reorient_vector_field_2d elif self.dim == 3: self.reorient_vector_field = vfu.reorient_vector_field_3d else: raise ValueError(f"SSD Metric not defined for dim. {self.dim}") if self.step_type == "gauss_newton": self.compute_step = self.compute_gauss_newton_step elif self.step_type == "demons": self.compute_step = self.compute_demons_step else: raise ValueError(f"Opt. step {self.step_type} not defined")
[docs] def initialize_iteration(self): r"""Prepares the metric to compute one displacement field iteration. Pre-computes the gradient of the input images to be used in the computation of the forward and backward steps. """ self.gradient_moving = np.empty( shape=self.moving_image.shape + (self.dim,), dtype=floating ) for i, grad in enumerate(gradient(self.moving_image)): self.gradient_moving[..., i] = grad # Convert static image's gradient field from voxel to physical space if self.moving_spacing is not None: self.gradient_moving /= self.moving_spacing if self.moving_direction is not None: self.reorient_vector_field(self.gradient_moving, self.moving_direction) self.gradient_static = np.empty( shape=self.static_image.shape + (self.dim,), dtype=floating ) for i, grad in enumerate(gradient(self.static_image)): self.gradient_static[..., i] = grad # Convert static image's gradient field from voxel to physical space if self.static_spacing is not None: self.gradient_static /= self.static_spacing if self.static_direction is not None: self.reorient_vector_field(self.gradient_static, self.static_direction)
[docs] def compute_forward(self): r"""Computes one step bringing the reference image towards the static. Computes the update displacement field to be used for registration of the moving image towards the static image """ return self.compute_step(forward_step=True)
[docs] def compute_backward(self): r"""Computes one step bringing the static image towards the moving. Computes the updated displacement field to be used for registration of the static image towards the moving image """ return self.compute_step(forward_step=False)
[docs] @warning_for_keywords() def compute_gauss_newton_step(self, *, forward_step=True): r"""Computes the Gauss-Newton energy minimization step Minimizes the linearized energy function (Newton step) defined by the sum of squared differences of corresponding pixels of the input images with respect to the displacement field. Parameters ---------- forward_step : boolean if True, computes the Newton step in the forward direction (warping the moving towards the static image). If False, computes the backward step (warping the static image to the moving image) Returns ------- displacement : array, shape = static_image.shape + (3,) if forward_step==True, the forward SSD Gauss-Newton step, else, the backward step """ reference_shape = self.static_image.shape if forward_step: gradient = self.gradient_static delta_field = self.static_image - self.moving_image else: gradient = self.gradient_moving delta_field = self.moving_image - self.static_image displacement = np.zeros(shape=reference_shape + (self.dim,), dtype=floating) if self.dim == 2: self.energy = v_cycle_2d( self.levels_below, self.inner_iter, delta_field, None, gradient, None, self.smooth, displacement, ) else: self.energy = v_cycle_3d( self.levels_below, self.inner_iter, delta_field, None, gradient, None, self.smooth, displacement, ) return displacement
[docs] @warning_for_keywords() def compute_demons_step(self, *, forward_step=True): r"""Demons step for SSD metric Computes the demons step proposed by :footcite:t:`Vercauteren2009` for the SSD metric. Parameters ---------- forward_step : boolean if True, computes the Demons step in the forward direction (warping the moving towards the static image). If False, computes the backward step (warping the static image to the moving image) Returns ------- displacement : array, shape (R, C, 2) or (S, R, C, 3) the Demons step References ---------- .. footbibliography:: """ sigma_reg_2 = np.sum(self.static_spacing**2) / self.dim if forward_step: gradient = self.gradient_static delta_field = self.static_image - self.moving_image else: gradient = self.gradient_moving delta_field = self.moving_image - self.static_image if self.dim == 2: step, self.energy = ssd.compute_ssd_demons_step_2d( delta_field, gradient, sigma_reg_2, None ) else: step, self.energy = ssd.compute_ssd_demons_step_3d( delta_field, gradient, sigma_reg_2, None ) for i in range(self.dim): step[..., i] = ndimage.gaussian_filter(step[..., i], self.smooth) return step
[docs] def get_energy(self): r"""The numerical value assigned by this metric to the current image pair Returns the Sum of Squared Differences (data term) energy computed at the largest iteration """ return self.energy
[docs] def free_iteration(self): r""" Nothing to free for the SSD metric """ pass
[docs] @warning_for_keywords() def v_cycle_2d( n, k, delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, *, depth=0, ): r"""Multi-resolution Gauss-Seidel solver using V-type cycles Multi-resolution Gauss-Seidel solver: solves the Gauss-Newton linear system by first filtering (GS-iterate) the current level, then solves for the residual at a coarser resolution and finally refines the solution at the current resolution. This scheme corresponds to the V-cycle proposed by :footcite:t:`Bruhn2005`. Parameters ---------- n : int number of levels of the multi-resolution algorithm (it will be called recursively until level n == 0) k : int the number of iterations at each multi-resolution level delta_field : array, shape (R, C) the difference between the static and moving image (the 'derivative w.r.t. time' in the optical flow model) sigma_sq_field : array, shape (R, C) the variance of the gray level value at each voxel, according to the EM model (for SSD, it is 1 for all voxels). Inf and 0 values are processed specially to support infinite and zero variance. gradient_field : array, shape (R, C, 2) the gradient of the moving image target : array, shape (R, C, 2) right-hand side of the linear system to be solved in the Weickert's multi-resolution algorithm lambda_param : float smoothness parameter, the larger its value the smoother the displacement field displacement : array, shape (R, C, 2) the displacement field to start the optimization from Returns ------- energy : the energy of the EM (or SSD if sigmafield[...]==1) metric at this iteration References ---------- .. footbibliography:: """ # pre-smoothing for _ in range(k): ssd.iterate_residual_displacement_field_ssd_2d( delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, ) if n == 0: energy = ssd.compute_energy_ssd_2d(delta_field) return energy # solve at coarser grid residual = None residual = ssd.compute_residual_displacement_field_ssd_2d( delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, residual, ) sub_residual = np.array(vfu.downsample_displacement_field_2d(residual)) del residual subsigma_sq_field = None if sigma_sq_field is not None: subsigma_sq_field = vfu.downsample_scalar_field_2d(sigma_sq_field) subdelta_field = vfu.downsample_scalar_field_2d(delta_field) subgradient_field = np.array(vfu.downsample_displacement_field_2d(gradient_field)) shape = np.array(displacement.shape).astype(np.int32) half_shape = ((shape[0] + 1) // 2, (shape[1] + 1) // 2, 2) sub_displacement = np.zeros(shape=half_shape, dtype=floating) sublambda_param = lambda_param * 0.25 v_cycle_2d( n - 1, k, subdelta_field, subsigma_sq_field, subgradient_field, sub_residual, sublambda_param, sub_displacement, depth=depth + 1, ) # displacement += np.array( # vfu.upsample_displacement_field(sub_displacement, shape)) displacement += vfu.resample_displacement_field_2d( sub_displacement, np.array([0.5, 0.5]), shape ) # post-smoothing for _ in range(k): ssd.iterate_residual_displacement_field_ssd_2d( delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, ) energy = ssd.compute_energy_ssd_2d(delta_field) return energy
[docs] @warning_for_keywords() def v_cycle_3d( n, k, delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, *, depth=0, ): r"""Multi-resolution Gauss-Seidel solver using V-type cycles Multi-resolution Gauss-Seidel solver: solves the linear system by first filtering (GS-iterate) the current level, then solves for the residual at a coarser resolution and finally refines the solution at the current resolution. This scheme corresponds to the V-cycle proposed by :footcite:t:`Bruhn2005`. Parameters ---------- n : int number of levels of the multi-resolution algorithm (it will be called recursively until level n == 0) k : int the number of iterations at each multi-resolution level delta_field : array, shape (S, R, C) the difference between the static and moving image (the 'derivative w.r.t. time' in the optical flow model) sigma_sq_field : array, shape (S, R, C) the variance of the gray level value at each voxel, according to the EM model (for SSD, it is 1 for all voxels). Inf and 0 values are processed specially to support infinite and zero variance. gradient_field : array, shape (S, R, C, 3) the gradient of the moving image target : array, shape (S, R, C, 3) right-hand side of the linear system to be solved in the Weickert's multi-resolution algorithm lambda_param : float smoothness parameter, the larger its value the smoother the displacement field displacement : array, shape (S, R, C, 3) the displacement field to start the optimization from Returns ------- energy : the energy of the EM (or SSD if sigmafield[...]==1) metric at this iteration References ---------- .. footbibliography:: """ # pre-smoothing for _ in range(k): ssd.iterate_residual_displacement_field_ssd_3d( delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, ) if n == 0: energy = ssd.compute_energy_ssd_3d(delta_field) return energy # solve at coarser grid residual = ssd.compute_residual_displacement_field_ssd_3d( delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, None, ) sub_residual = np.array(vfu.downsample_displacement_field_3d(residual)) del residual subsigma_sq_field = None if sigma_sq_field is not None: subsigma_sq_field = vfu.downsample_scalar_field_3d(sigma_sq_field) subdelta_field = vfu.downsample_scalar_field_3d(delta_field) subgradient_field = np.array(vfu.downsample_displacement_field_3d(gradient_field)) shape = np.array(displacement.shape).astype(np.int32) sub_displacement = np.zeros( shape=((shape[0] + 1) // 2, (shape[1] + 1) // 2, (shape[2] + 1) // 2, 3), dtype=floating, ) sublambda_param = lambda_param * 0.25 v_cycle_3d( n - 1, k, subdelta_field, subsigma_sq_field, subgradient_field, sub_residual, sublambda_param, sub_displacement, depth=depth + 1, ) del subdelta_field del subsigma_sq_field del subgradient_field del sub_residual displacement += vfu.resample_displacement_field_3d( sub_displacement, 0.5 * np.ones(3), shape ) del sub_displacement # post-smoothing for _ in range(k): ssd.iterate_residual_displacement_field_ssd_3d( delta_field, sigma_sq_field, gradient_field, target, lambda_param, displacement, ) energy = ssd.compute_energy_ssd_3d(delta_field) return energy