"""Implementation of various restoration functions.""" import numpy as np from scipy.signal import convolve from .._shared.utils import _supported_float_type from . import uft def wiener(image, psf, balance, reg=None, is_real=True, clip=True): r"""Restore image using Wiener–Hunt deconvolution. Wiener–Hunt deconvolution is a restoration method which follows a Bayesian approach [1]_. Parameters ---------- image : (N1, N2, ..., ND) ndarray Degraded image. psf : ndarray Point spread function (PSF). Assumed to be the impulse response (input image space) if the data type is real, or the transfer function (Fourier or frequency space) if the data type is complex. There is no constraint on the shape of the impulse response. The transfer function though must be of shape `(N1, N2, ..., ND)` if `is_real is True`, `(N1, N2, ..., ND // 2 + 1)` otherwise (see :func:`numpy.fft.rfftn`). balance : float Regularization parameter. Denoted by :math:`\lambda`: in the Notes section below, its value lets you balance data adequacy (improving frequency restoration) with respect to prior adequacy (reducing frequency restoration and avoiding noise artifacts). A larger value for this parameter favors the regularization/prior. reg : ndarray, optional Regularization operator. Laplacian by default. It can be an impulse response or a transfer function, as for the PSF. Shape constraints are the same as for `psf`. is_real : bool, optional True by default. Specify if `psf` and `reg` are provided over just half the frequency space (thanks to the redundancy of the Fourier transform for real signals). Applies only if `psf` and/or `reg` are provided as transfer functions. See ``uft`` module and :func:`np.fft.rfftn`. clip : bool, optional True by default. If True, pixel values of the deconvolved image (which is the return value) above 1 (resp. below -1) are clipped to 1 (resp. to -1). Be careful to set `clip=False` if you do not want this clipping and/or if your data range is not [0, 1] or [-1,1]. Returns ------- im_deconv : (N1, N2, ..., ND) ndarray The deconvolved image. Examples -------- >>> import skimage as ski >>> import scipy as sp >>> img = ski.color.rgb2gray(ski.data.astronaut()) >>> psf = np.ones((5, 5)) / 25 >>> img = sp.signal.convolve2d(img, psf, 'same') >>> rng = np.random.default_rng() >>> img += 0.1 * img.std() * rng.standard_normal(img.shape) >>> deconvolved_img = ski.restoration.wiener(img, psf, 0.1) Notes ----- This function applies the Wiener filter to a noisy (degraded) image by an impulse response (or PSF). If the data model is .. math:: y = Hx + n where :math:`n` is noise, :math:`H` the PSF, and :math:`x` the unknown original image, the Wiener filter is .. math:: \hat x = F^\dagger \left( |\Lambda_H|^2 + \lambda |\Lambda_D|^2 \right)^{-1} \Lambda_H^\dagger F y where :math:`F` and :math:`F^\dagger` are the Fourier and inverse Fourier transforms respectively, :math:`\Lambda_H` the transfer function (or the Fourier transform of the PSF, see [2]_), and :math:`\Lambda_D` the regularization operator, which is a filter penalizing the restored image frequencies (Laplacian by default, that is, penalization of high frequencies). The parameter :math:`\lambda` tunes the balance between data (which tends to increase high frequencies, even those coming from noise) and regularization/prior (which tends to avoid noise artifacts). These methods are then specific to a prior model. Consequently, the application or the true image nature must correspond to the prior model. By default, the prior model (Laplacian) introduces image smoothness or pixel correlation. It can also be interpreted as high-frequency penalization to compensate for the instability of the solution with respect to the data (sometimes called noise amplification or "explosive" solution). Finally, the use of Fourier space implies a circulant property of :math:`H`, see [2]_. References ---------- .. [1] François Orieux, Jean-François Giovannelli, and Thomas Rodet, "Bayesian estimation of regularization and point spread function parameters for Wiener–Hunt deconvolution", J. Opt. Soc. Am. A 27, 1593–1607 (2010) https://www.osapublishing.org/josaa/abstract.cfm?URI=josaa-27-7-1593 https://hal.archives-ouvertes.fr/hal-00674508 .. [2] B. R. Hunt "A matrix theory proof of the discrete convolution theorem", IEEE Trans. on Audio and Electroacoustics, vol. au-19, no. 4, pp. 285–288, dec. 1971 """ if reg is None: reg, _ = uft.laplacian(image.ndim, image.shape, is_real=is_real) if not np.iscomplexobj(reg): reg = uft.ir2tf(reg, image.shape, is_real=is_real) float_type = _supported_float_type(image.dtype) image = image.astype(float_type, copy=False) psf = psf.real.astype(float_type, copy=False) reg = reg.real.astype(float_type, copy=False) if psf.shape != reg.shape: trans_func = uft.ir2tf(psf, image.shape, is_real=is_real) else: trans_func = psf wiener_filter = np.conj(trans_func) / ( np.abs(trans_func) ** 2 + balance * np.abs(reg) ** 2 ) if is_real: deconv = uft.uirfftn(wiener_filter * uft.urfftn(image), shape=image.shape) else: deconv = uft.uifftn(wiener_filter * uft.ufftn(image)) if clip: deconv[deconv > 1] = 1 deconv[deconv < -1] = -1 return deconv def unsupervised_wiener( image, psf, reg=None, user_params=None, is_real=True, clip=True, *, rng=None ): """Unsupervised Wiener-Hunt deconvolution. Return the deconvolution with a Wiener-Hunt approach, where the hyperparameters are automatically estimated. The algorithm is a stochastic iterative process (Gibbs sampler) described in the reference below. See also ``wiener`` function. Parameters ---------- image : (M, N) ndarray The input degraded image. psf : ndarray The impulse response (input image's space) or the transfer function (Fourier space). Both are accepted. The transfer function is automatically recognized as being complex (``np.iscomplexobj(psf)``). reg : ndarray, optional The regularisation operator. The Laplacian by default. It can be an impulse response or a transfer function, as for the psf. user_params : dict, optional Dictionary of parameters for the Gibbs sampler. Accepted keys are: threshold : float The stopping criterion: the norm of the difference between to successive approximated solution (empirical mean of object samples, see Notes section). 1e-4 by default. burnin : int The number of sample to ignore to start computation of the mean. 15 by default. min_num_iter : int The minimum number of iterations. 30 by default. max_num_iter : int The maximum number of iterations if ``threshold`` is not satisfied. 200 by default. callback : callable A user provided callable to which is passed, if the function exists, the current image sample for whatever purpose. The user can store the sample, or compute other moments than the mean. It has no influence on the algorithm execution and is only for inspection. clip : bool, optional True by default. If true, pixel values of the result above 1 or under -1 are thresholded for skimage pipeline compatibility. rng : {`numpy.random.Generator`, int}, optional Pseudo-random number generator. By default, a PCG64 generator is used (see :func:`numpy.random.default_rng`). If `rng` is an int, it is used to seed the generator. .. versionadded:: 0.19 Returns ------- x_postmean : (M, N) ndarray The deconvolved image (the posterior mean). chains : dict The keys ``noise`` and ``prior`` contain the chain list of noise and prior precision respectively. Examples -------- >>> from skimage import color, data, restoration >>> img = color.rgb2gray(data.astronaut()) >>> from scipy.signal import convolve2d >>> psf = np.ones((5, 5)) / 25 >>> img = convolve2d(img, psf, 'same') >>> rng = np.random.default_rng() >>> img += 0.1 * img.std() * rng.standard_normal(img.shape) >>> deconvolved_img = restoration.unsupervised_wiener(img, psf) Notes ----- The estimated image is design as the posterior mean of a probability law (from a Bayesian analysis). The mean is defined as a sum over all the possible images weighted by their respective probability. Given the size of the problem, the exact sum is not tractable. This algorithm use of MCMC to draw image under the posterior law. The practical idea is to only draw highly probable images since they have the biggest contribution to the mean. At the opposite, the less probable images are drawn less often since their contribution is low. Finally, the empirical mean of these samples give us an estimation of the mean, and an exact computation with an infinite sample set. References ---------- .. [1] François Orieux, Jean-François Giovannelli, and Thomas Rodet, "Bayesian estimation of regularization and point spread function parameters for Wiener-Hunt deconvolution", J. Opt. Soc. Am. A 27, 1593-1607 (2010) https://www.osapublishing.org/josaa/abstract.cfm?URI=josaa-27-7-1593 https://hal.archives-ouvertes.fr/hal-00674508 """ params = { 'threshold': 1e-4, 'max_num_iter': 200, 'min_num_iter': 30, 'burnin': 15, 'callback': None, } params.update(user_params or {}) if reg is None: reg, _ = uft.laplacian(image.ndim, image.shape, is_real=is_real) if not np.iscomplexobj(reg): reg = uft.ir2tf(reg, image.shape, is_real=is_real) float_type = _supported_float_type(image.dtype) image = image.astype(float_type, copy=False) psf = psf.real.astype(float_type, copy=False) reg = reg.real.astype(float_type, copy=False) if psf.shape != reg.shape: trans_fct = uft.ir2tf(psf, image.shape, is_real=is_real) else: trans_fct = psf # The mean of the object x_postmean = np.zeros(trans_fct.shape, dtype=float_type) # The previous computed mean in the iterative loop prev_x_postmean = np.zeros(trans_fct.shape, dtype=float_type) # Difference between two successive mean delta = np.nan # Initial state of the chain gn_chain, gx_chain = [1], [1] # The correlation of the object in Fourier space (if size is big, # this can reduce computation time in the loop) areg2 = np.abs(reg) ** 2 atf2 = np.abs(trans_fct) ** 2 # The Fourier transform may change the image.size attribute, so we # store it. if is_real: data_spectrum = uft.urfft2(image) else: data_spectrum = uft.ufft2(image) rng = np.random.default_rng(rng) # Gibbs sampling for iteration in range(params['max_num_iter']): # Sample of Eq. 27 p(circX^k | gn^k-1, gx^k-1, y). # weighting (correlation in direct space) precision = gn_chain[-1] * atf2 + gx_chain[-1] * areg2 # Eq. 29 # Note: Use astype instead of dtype argument to standard_normal to get # similar random values across precisions, as needed for # reference data used by test_unsupervised_wiener. _rand1 = rng.standard_normal(data_spectrum.shape) _rand1 = _rand1.astype(float_type, copy=False) _rand2 = rng.standard_normal(data_spectrum.shape) _rand2 = _rand2.astype(float_type, copy=False) excursion = np.sqrt(0.5 / precision) * (_rand1 + 1j * _rand2) # mean Eq. 30 (RLS for fixed gn, gamma0 and gamma1 ...) wiener_filter = gn_chain[-1] * np.conj(trans_fct) / precision # sample of X in Fourier space x_sample = wiener_filter * data_spectrum + excursion if params['callback']: params['callback'](x_sample) # sample of Eq. 31 p(gn | x^k, gx^k, y) gn_chain.append( rng.gamma( image.size / 2, 2 / uft.image_quad_norm(data_spectrum - x_sample * trans_fct), ) ) # sample of Eq. 31 p(gx | x^k, gn^k-1, y) gx_chain.append( rng.gamma((image.size - 1) / 2, 2 / uft.image_quad_norm(x_sample * reg)) ) # current empirical average if iteration > params['burnin']: x_postmean = prev_x_postmean + x_sample if iteration > (params['burnin'] + 1): current = x_postmean / (iteration - params['burnin']) previous = prev_x_postmean / (iteration - params['burnin'] - 1) delta = ( np.sum(np.abs(current - previous)) / np.sum(np.abs(x_postmean)) / (iteration - params['burnin']) ) prev_x_postmean = x_postmean # stop of the algorithm if (iteration > params['min_num_iter']) and (delta < params['threshold']): break # Empirical average \approx POSTMEAN Eq. 44 x_postmean = x_postmean / (iteration - params['burnin']) if is_real: x_postmean = uft.uirfft2(x_postmean, shape=image.shape) else: x_postmean = uft.uifft2(x_postmean) if clip: x_postmean[x_postmean > 1] = 1 x_postmean[x_postmean < -1] = -1 return (x_postmean, {'noise': gn_chain, 'prior': gx_chain}) def richardson_lucy(image, psf, num_iter=50, clip=True, filter_epsilon=None): """Richardson-Lucy deconvolution. Parameters ---------- image : ([P, ]M, N) ndarray Input degraded image (can be n-dimensional). If you keep the default `clip=True` parameter, you may want to normalize the image so that its values fall in the [-1, 1] interval to avoid information loss. psf : ndarray The point spread function. num_iter : int, optional Number of iterations. This parameter plays the role of regularisation. clip : bool, optional True by default. If true, pixel value of the result above 1 or under -1 are thresholded for skimage pipeline compatibility. filter_epsilon : float, optional Value below which intermediate results become 0 to avoid division by small numbers. Returns ------- im_deconv : ndarray The deconvolved image. Examples -------- >>> from skimage import img_as_float, data, restoration >>> camera = img_as_float(data.camera()) >>> from scipy.signal import convolve2d >>> psf = np.ones((5, 5)) / 25 >>> camera = convolve2d(camera, psf, 'same') >>> rng = np.random.default_rng() >>> camera += 0.1 * camera.std() * rng.standard_normal(camera.shape) >>> deconvolved = restoration.richardson_lucy(camera, psf, 5) References ---------- .. [1] https://en.wikipedia.org/wiki/Richardson%E2%80%93Lucy_deconvolution """ float_type = _supported_float_type(image.dtype) image = image.astype(float_type, copy=False) psf = psf.astype(float_type, copy=False) im_deconv = np.full(image.shape, 0.5, dtype=float_type) psf_mirror = np.flip(psf) # Small regularization parameter used to avoid 0 divisions eps = 1e-12 for _ in range(num_iter): conv = convolve(im_deconv, psf, mode='same') + eps if filter_epsilon: relative_blur = np.where(conv < filter_epsilon, 0, image / conv) else: relative_blur = image / conv im_deconv *= convolve(relative_blur, psf_mirror, mode='same') if clip: im_deconv[im_deconv > 1] = 1 im_deconv[im_deconv < -1] = -1 return im_deconv