Source code for bob.bio.face.extractor.LGBPHS

#!/usr/bin/env python
# vim: set fileencoding=utf-8 :
# Manuel Guenther <Manuel.Guenther@idiap.ch>

import bob.ip.gabor
import bob.ip.base

import numpy
import math

from bob.bio.base.extractor import Extractor

class LGBPHS (Extractor):
  """Extracts *Local Gabor Binary Pattern Histogram Sequences* (LGBPHS) [ZSG+05]_ from the images, using functionality from :ref:`bob.ip.base <bob.ip.base>` and :ref:`bob.ip.gabor <bob.ip.gabor>`.

  The block size and the overlap of the blocks can be varied, as well as the parameters of the Gabor wavelet (:py:class:`bob.ip.gabor.Transform`) and the LBP extractor (:py:class:`bob.ip.base.LBP`).

  **Parameters:**

  block_size : int or (int, int)
    The size of the blocks that will be extracted.
    This parameter might be either a single integral value, or a pair ``(block_height, block_width)`` of integral values.

  block_overlap : int or (int, int)
    The overlap of the blocks in vertical and horizontal direction.
    This parameter might be either a single integral value, or a pair ``(block_overlap_y, block_overlap_x)`` of integral values.
    It needs to be smaller than the ``block_size``.

  gabor_directions, gabor_scales, gabor_sigma, gabor_maximum_frequency, gabor_frequency_step, gabor_power_of_k, gabor_dc_free
    The parameters of the Gabor wavelet family, with its default values set as given in [WFK97]_.
    Please refer to :py:class:`bob.ip.gabor.Transform` for the documentation of these values.

  use_gabor_phases : bool
    Extract also the Gabor phases (inline) and not only the absolute values.
    In this case, Extended LGBPHS features [ZSQ+09]_ will be extracted.

  lbp_radius, lbp_neighbor_count, lbp_uniform, lbp_circular, lbp_rotation_invariant, lbp_compare_to_average, lbp_add_average
    The parameters of the LBP.
    Please see :py:class:`bob.ip.base.LBP` for the documentation of these values.

    .. note::
       The default values are as given in [ZSG+05]_ (the values of [ZSQ+09]_ might differ).

  sparse_histogram : bool
    If specified, the histograms will be handled in a sparse way.
    This reduces the size of the extracted features, but the computation will take longer.

    .. note::
       Sparse histograms are only supported, when ``split_histogram = None``.

  split_histogram : one of ``('blocks', 'wavelets', 'both')`` or ``None``
    Defines, how the histogram sequence is split.
    This could be interesting, if the histograms should be used in another way as simply concatenating them into a single histogram sequence (the default).
  """

  def __init__(
      self,
      # Block setup
      block_size,    # one or two parameters for block size
      block_overlap = 0, # one or two parameters for block overlap
      # Gabor parameters
      gabor_directions = 8,
      gabor_scales = 5,
      gabor_sigma = 2. * math.pi,
      gabor_maximum_frequency = math.pi / 2.,
      gabor_frequency_step = math.sqrt(.5),
      gabor_power_of_k = 0,
      gabor_dc_free = True,
      use_gabor_phases = False,
      # LBP parameters
      lbp_radius = 2,
      lbp_neighbor_count = 8,
      lbp_uniform = True,
      lbp_circular = True,
      lbp_rotation_invariant = False,
      lbp_compare_to_average = False,
      lbp_add_average = False,
      # histogram options
      sparse_histogram = False,
      split_histogram = None
  ):
    # call base class constructor
    Extractor.__init__(
        self,

        block_size = block_size,
        block_overlap = block_overlap,
        gabor_directions = gabor_directions,
        gabor_scales = gabor_scales,
        gabor_sigma = gabor_sigma,
        gabor_maximum_frequency = gabor_maximum_frequency,
        gabor_frequency_step = gabor_frequency_step,
        gabor_power_of_k = gabor_power_of_k,
        gabor_dc_free = gabor_dc_free,
        use_gabor_phases = use_gabor_phases,
        lbp_radius = lbp_radius,
        lbp_neighbor_count = lbp_neighbor_count,
        lbp_uniform = lbp_uniform,
        lbp_circular = lbp_circular,
        lbp_rotation_invariant = lbp_rotation_invariant,
        lbp_compare_to_average = lbp_compare_to_average,
        lbp_add_average = lbp_add_average,
        sparse_histogram = sparse_histogram,
        split_histogram = split_histogram
    )

    # block parameters
    self.block_size = block_size if isinstance(block_size, (tuple, list)) else (block_size, block_size)
    self.block_overlap = block_overlap if isinstance(block_overlap, (tuple, list)) else (block_overlap, block_overlap)
    if self.block_size[0] < self.block_overlap[0] or self.block_size[1] < self.block_overlap[1]:
      raise ValueError("The overlap is bigger than the block size. This won't work. Please check your setup!")

    # Gabor wavelet transform class
    self.gwt = bob.ip.gabor.Transform(
        number_of_scales = gabor_scales,
        number_of_directions = gabor_directions,
        sigma = gabor_sigma,
        k_max = gabor_maximum_frequency,
        k_fac = gabor_frequency_step,
        power_of_k = gabor_power_of_k,
        dc_free = gabor_dc_free
    )
    self.trafo_image = None
    self.use_phases = use_gabor_phases

    self.lbp = bob.ip.base.LBP(
        neighbors = lbp_neighbor_count,
        radius = float(lbp_radius),
        circular = lbp_circular,
        to_average = lbp_compare_to_average,
        add_average_bit = lbp_add_average,
        uniform = lbp_uniform,
        rotation_invariant = lbp_rotation_invariant,
        border_handling = 'wrap'
    )

    self.split = split_histogram
    self.sparse = sparse_histogram
    if self.sparse and self.split:
      raise ValueError("Sparse histograms cannot be split! Check your setup!")


  def _fill(self, lgbphs_array, lgbphs_blocks, j):
    """Copies the given array into the given blocks"""
    # fill array in the desired shape
    if self.split is None:
      start = j * self.n_bins * self.n_blocks
      for b in range(self.n_blocks):
        lgbphs_array[start + b * self.n_bins : start + (b+1) * self.n_bins] = lgbphs_blocks[b][:]
    elif self.split == 'blocks':
      for b in range(self.n_blocks):
        lgbphs_array[b, j * self.n_bins : (j+1) * self.n_bins] = lgbphs_blocks[b][:]
    elif self.split == 'wavelets':
      for b in range(self.n_blocks):
        lgbphs_array[j, b * self.n_bins : (b+1) * self.n_bins] = lgbphs_blocks[b][:]
    elif self.split == 'both':
      for b in range(self.n_blocks):
        lgbphs_array[j * self.n_blocks + b, 0 : self.n_bins] = lgbphs_blocks[b][:]

  def _sparsify(self, array):
    """This function generates a sparse histogram from a non-sparse one."""
    if not self.sparse:
      return array
    if len(array.shape) == 2 and array.shape[0] == 2:
      # already sparse
      return array
    assert len(array.shape) == 1
    indices = []
    values = []
    for i in range(array.shape[0]):
      if array[i] != 0.:
        indices.append(i)
        values.append(array[i])
    return numpy.array([indices, values], dtype = numpy.float64)


  def __call__(self, image):
    """__call__(image) -> feature

    Extracts the local Gabor binary pattern histogram sequence from the given image.

    **Parameters:**

    image : 2D :py:class:`numpy.ndarray` (floats)
      The image to extract the features from.

    **Returns:**

    feature : 2D or 3D :py:class:`numpy.ndarray` (floats)
      The list of Gabor jets extracted from the image.
      The 2D location of the jet's nodes is not returned.
    """
    """"""
    assert image.ndim == 2
    assert isinstance(image, numpy.ndarray)
    assert image.dtype == numpy.float64

    # perform GWT on image
    if self.trafo_image is None or self.trafo_image.shape[1:3] != image.shape:
      # create trafo image
      self.trafo_image = numpy.ndarray((self.gwt.number_of_wavelets, image.shape[0], image.shape[1]), numpy.complex128)

    # perform Gabor wavelet transform
    self.gwt.transform(image, self.trafo_image)

    jet_length = self.gwt.number_of_wavelets * (2 if self.use_phases else 1)

    lgbphs_array = None
    # iterate through the layers of the trafo image
    for j in range(self.gwt.number_of_wavelets):
      # compute absolute part of complex response
      abs_image = numpy.abs(self.trafo_image[j])
      # Computes LBP histograms
      abs_blocks = bob.ip.base.lbphs(abs_image, self.lbp, self.block_size, self.block_overlap)

      # Converts to Blitz array (of different dimensionalities)
      self.n_bins = abs_blocks.shape[1]
      self.n_blocks = abs_blocks.shape[0]

      if self.split is None:
        shape = (self.n_blocks * self.n_bins * jet_length,)
      elif self.split == 'blocks':
        shape = (self.n_blocks, self.n_bins * jet_length)
      elif self.split == 'wavelets':
        shape = (jet_length, self.n_bins * self.n_blocks)
      elif self.split == 'both':
        shape = (jet_length * self.n_blocks, self.n_bins)
      else:
        raise ValueError("The split parameter must be one of ['blocks', 'wavelets', 'both'] or None")

      # create new array if not done yet
      if lgbphs_array is None:
        lgbphs_array = numpy.ndarray(shape, 'float64')

      # fill the array with the absolute values of the Gabor wavelet transform
      self._fill(lgbphs_array, abs_blocks, j)

      if self.use_phases:
        # compute phase part of complex response
        phase_image = numpy.angle(self.trafo_image[j])
        # Computes LBP histograms
        phase_blocks = bob.ip.base.lbphs(phase_image, self.lbp, self.block_size, self.block_overlap)
        # fill the array with the phases at the end of the blocks
        self._fill(lgbphs_array, phase_blocks, j + self.gwt.number_of_wavelets)


    # return the concatenated list of all histograms
    return self._sparsify(lgbphs_array)

  # re-define the train function to get it non-documented
[docs] def train(*args,**kwargs) : raise NotImplementedError("This function is not implemented and should not be called.")
[docs] def load(*args,**kwargs) : pass