""" Augmenters that perform contrast changes. List of augmenters: * :class:`GammaContrast` * :class:`SigmoidContrast` * :class:`LogContrast` * :class:`LinearContrast` * :class:`AllChannelsHistogramEqualization` * :class:`HistogramEqualization` * :class:`AllChannelsCLAHE` * :class:`CLAHE` """ from __future__ import print_function, division, absolute_import import numpy as np import six.moves as sm import skimage.exposure as ski_exposure import cv2 import imgaug as ia from imgaug.imgaug import _normalize_cv2_input_arr_ from . import meta from . import color as color_lib from .. import parameters as iap from .. import dtypes as iadt from ..augmentables.batches import _BatchInAugmentation class _ContrastFuncWrapper(meta.Augmenter): def __init__(self, func, params1d, per_channel, dtypes_allowed=None, dtypes_disallowed=None, seed=None, name=None, random_state="deprecated", deterministic="deprecated"): super(_ContrastFuncWrapper, self).__init__( seed=seed, name=name, random_state=random_state, deterministic=deterministic) self.func = func self.params1d = params1d self.per_channel = iap.handle_probability_param(per_channel, "per_channel") self.dtypes_allowed = dtypes_allowed self.dtypes_disallowed = dtypes_disallowed # Added in 0.4.0. def _augment_batch_(self, batch, random_state, parents, hooks): if batch.images is None: return batch images = batch.images if self.dtypes_allowed is not None: iadt.gate_dtypes(images, allowed=self.dtypes_allowed, disallowed=self.dtypes_disallowed, augmenter=self) nb_images = len(images) rss = random_state.duplicate(1+nb_images) per_channel = self.per_channel.draw_samples((nb_images,), random_state=rss[0]) gen = enumerate(zip(images, per_channel, rss[1:])) for i, (image, per_channel_i, rs) in gen: nb_channels = 1 if per_channel_i <= 0.5 else image.shape[2] # TODO improve efficiency by sampling once samples_i = [ param.draw_samples((nb_channels,), random_state=rs) for param in self.params1d] if per_channel_i > 0.5: input_dtype = image.dtype # TODO This was previously a cast of image to float64. Do the # adjust_* functions return float64? result = [] for c in sm.xrange(nb_channels): samples_i_c = [sample_i[c] for sample_i in samples_i] args = tuple([image[..., c]] + samples_i_c) result.append(self.func(*args)) image_aug = np.stack(result, axis=-1) image_aug = image_aug.astype(input_dtype) else: # don't use something like samples_i[...][0] here, because # that returns python scalars and is slightly less accurate # than keeping the numpy values args = tuple([image] + samples_i) image_aug = self.func(*args) batch.images[i] = image_aug return batch def get_parameters(self): """See :func:`~imgaug.augmenters.meta.Augmenter.get_parameters`.""" return self.params1d # TODO quite similar to the other adjust_contrast_*() functions, make DRY def adjust_contrast_gamma(arr, gamma): """ Adjust image contrast by scaling pixel values to ``255*((v/255)**gamma)``. **Supported dtypes**: * ``uint8``: yes; fully tested (1) (2) (3) * ``uint16``: yes; tested (2) (3) * ``uint32``: yes; tested (2) (3) * ``uint64``: yes; tested (2) (3) (4) * ``int8``: limited; tested (2) (3) (5) * ``int16``: limited; tested (2) (3) (5) * ``int32``: limited; tested (2) (3) (5) * ``int64``: limited; tested (2) (3) (4) (5) * ``float16``: limited; tested (5) * ``float32``: limited; tested (5) * ``float64``: limited; tested (5) * ``float128``: no (6) * ``bool``: no (7) - (1) Handled by ``cv2``. Other dtypes are handled by ``skimage``. - (2) Normalization is done as ``I_ij/max``, where ``max`` is the maximum value of the dtype, e.g. 255 for ``uint8``. The normalization is reversed afterwards, e.g. ``result*255`` for ``uint8``. - (3) Integer-like values are not rounded after applying the contrast adjustment equation (before inverting the normalization to ``[0.0, 1.0]`` space), i.e. projection from continuous space to discrete happens according to floor function. - (4) Note that scikit-image doc says that integers are converted to ``float64`` values before applying the contrast normalization method. This might lead to inaccuracies for large 64bit integer values. Tests showed no indication of that happening though. - (5) Must not contain negative values. Values >=0 are fully supported. - (6) Leads to error in scikit-image. - (7) Does not make sense for contrast adjustments. Parameters ---------- arr : numpy.ndarray Array for which to adjust the contrast. Dtype ``uint8`` is fastest. gamma : number Exponent for the contrast adjustment. Higher values darken the image. Returns ------- numpy.ndarray Array with adjusted contrast. """ if arr.size == 0: return np.copy(arr) # int8 is also possible according to docs # https://docs.opencv.org/3.0-beta/modules/core/doc/operations_on_arrays.html#cv2.LUT , # but here it seemed like `d` was 0 for CV_8S, causing that to fail if arr.dtype.name == "uint8": min_value, _center_value, max_value = \ iadt.get_value_range_of_dtype(arr.dtype) dynamic_range = max_value - min_value value_range = np.linspace(0, 1.0, num=dynamic_range+1, dtype=np.float32) # 255 * ((I_ij/255)**gamma) # using np.float32(.) here still works when the input is a numpy array # of size 1 table = (min_value + (value_range ** np.float32(gamma)) * dynamic_range) table = np.clip(table, min_value, max_value).astype(arr.dtype) arr_aug = ia.apply_lut(arr, table) return arr_aug return ski_exposure.adjust_gamma(arr, gamma) # TODO quite similar to the other adjust_contrast_*() functions, make DRY def adjust_contrast_sigmoid(arr, gain, cutoff): """ Adjust image contrast to ``255*1/(1+exp(gain*(cutoff-I_ij/255)))``. **Supported dtypes**: * ``uint8``: yes; fully tested (1) (2) (3) * ``uint16``: yes; tested (2) (3) * ``uint32``: yes; tested (2) (3) * ``uint64``: yes; tested (2) (3) (4) * ``int8``: limited; tested (2) (3) (5) * ``int16``: limited; tested (2) (3) (5) * ``int32``: limited; tested (2) (3) (5) * ``int64``: limited; tested (2) (3) (4) (5) * ``float16``: limited; tested (5) * ``float32``: limited; tested (5) * ``float64``: limited; tested (5) * ``float128``: no (6) * ``bool``: no (7) - (1) Handled by ``cv2``. Other dtypes are handled by ``skimage``. - (2) Normalization is done as ``I_ij/max``, where ``max`` is the maximum value of the dtype, e.g. 255 for ``uint8``. The normalization is reversed afterwards, e.g. ``result*255`` for ``uint8``. - (3) Integer-like values are not rounded after applying the contrast adjustment equation before inverting the normalization to ``[0.0, 1.0]`` space), i.e. projection from continuous space to discrete happens according to floor function. - (4) Note that scikit-image doc says that integers are converted to ``float64`` values before applying the contrast normalization method. This might lead to inaccuracies for large 64bit integer values. Tests showed no indication of that happening though. - (5) Must not contain negative values. Values >=0 are fully supported. - (6) Leads to error in scikit-image. - (7) Does not make sense for contrast adjustments. Parameters ---------- arr : numpy.ndarray Array for which to adjust the contrast. Dtype ``uint8`` is fastest. gain : number Multiplier for the sigmoid function's output. Higher values lead to quicker changes from dark to light pixels. cutoff : number Cutoff that shifts the sigmoid function in horizontal direction. Higher values mean that the switch from dark to light pixels happens later, i.e. the pixels will remain darker. Returns ------- numpy.ndarray Array with adjusted contrast. """ if arr.size == 0: return np.copy(arr) # int8 is also possible according to docs # https://docs.opencv.org/3.0-beta/modules/core/doc/operations_on_arrays.html#cv2.LUT , # but here it seemed like `d` was 0 for CV_8S, causing that to fail if arr.dtype.name == "uint8": min_value, _center_value, max_value = \ iadt.get_value_range_of_dtype(arr.dtype) dynamic_range = max_value - min_value value_range = np.linspace(0, 1.0, num=dynamic_range+1, dtype=np.float32) # 255 * 1/(1 + exp(gain*(cutoff - I_ij/255))) # using np.float32(.) here still works when the input is a numpy array # of size 1 gain = np.float32(gain) cutoff = np.float32(cutoff) table = (min_value + dynamic_range * 1/(1 + np.exp(gain * (cutoff - value_range)))) table = np.clip(table, min_value, max_value).astype(arr.dtype) arr_aug = ia.apply_lut(arr, table) return arr_aug return ski_exposure.adjust_sigmoid(arr, cutoff=cutoff, gain=gain) # TODO quite similar to the other adjust_contrast_*() functions, make DRY # TODO add dtype gating def adjust_contrast_log(arr, gain): """ Adjust image contrast by scaling pixels to ``255*gain*log_2(1+v/255)``. **Supported dtypes**: * ``uint8``: yes; fully tested (1) (2) (3) * ``uint16``: yes; tested (2) (3) * ``uint32``: no; tested (2) (3) (8) * ``uint64``: no; tested (2) (3) (4) (8) * ``int8``: limited; tested (2) (3) (5) * ``int16``: limited; tested (2) (3) (5) * ``int32``: no; tested (2) (3) (5) (8) * ``int64``: no; tested (2) (3) (4) (5) (8) * ``float16``: limited; tested (5) * ``float32``: limited; tested (5) * ``float64``: limited; tested (5) * ``float128``: no (6) * ``bool``: no (7) - (1) Handled by ``cv2``. Other dtypes are handled by ``skimage``. - (2) Normalization is done as ``I_ij/max``, where ``max`` is the maximum value of the dtype, e.g. 255 for ``uint8``. The normalization is reversed afterwards, e.g. ``result*255`` for ``uint8``. - (3) Integer-like values are not rounded after applying the contrast adjustment equation (before inverting the normalization to ``[0.0, 1.0]`` space), i.e. projection from continuous space to discrete happens according to floor function. - (4) Note that scikit-image doc says that integers are converted to ``float64`` values before applying the contrast normalization method. This might lead to inaccuracies for large 64bit integer values. Tests showed no indication of that happening though. - (5) Must not contain negative values. Values >=0 are fully supported. - (6) Leads to error in scikit-image. - (7) Does not make sense for contrast adjustments. - (8) No longer supported since numpy 1.17. Previously: 'yes' for ``uint32``, ``uint64``; 'limited' for ``int32``, ``int64``. Parameters ---------- arr : numpy.ndarray Array for which to adjust the contrast. Dtype ``uint8`` is fastest. gain : number Multiplier for the logarithm result. Values around 1.0 lead to a contrast-adjusted images. Values above 1.0 quickly lead to partially broken images due to exceeding the datatype's value range. Returns ------- numpy.ndarray Array with adjusted contrast. """ if arr.size == 0: return np.copy(arr) # int8 is also possible according to docs # https://docs.opencv.org/3.0-beta/modules/core/doc/operations_on_arrays.html#cv2.LUT , # but here it seemed like `d` was 0 for CV_8S, causing that to fail if arr.dtype.name == "uint8": min_value, _center_value, max_value = \ iadt.get_value_range_of_dtype(arr.dtype) dynamic_range = max_value - min_value value_range = np.linspace(0, 1.0, num=dynamic_range+1, dtype=np.float32) # 255 * 1/(1 + exp(gain*(cutoff - I_ij/255))) # using np.float32(.) here still works when the input is a numpy array # of size 1 gain = np.float32(gain) table = min_value + dynamic_range * gain * np.log2(1 + value_range) table = np.clip(table, min_value, max_value).astype(arr.dtype) arr_aug = ia.apply_lut(arr, table) return arr_aug return ski_exposure.adjust_log(arr, gain=gain) # TODO quite similar to the other adjust_contrast_*() functions, make DRY def adjust_contrast_linear(arr, alpha): """Adjust contrast by scaling each pixel to ``127 + alpha*(v-127)``. **Supported dtypes**: * ``uint8``: yes; fully tested (1) (2) * ``uint16``: yes; tested (2) * ``uint32``: yes; tested (2) * ``uint64``: no (3) * ``int8``: yes; tested (2) * ``int16``: yes; tested (2) * ``int32``: yes; tested (2) * ``int64``: no (2) * ``float16``: yes; tested (2) * ``float32``: yes; tested (2) * ``float64``: yes; tested (2) * ``float128``: no (2) * ``bool``: no (4) - (1) Handled by ``cv2``. Other dtypes are handled by raw ``numpy``. - (2) Only tested for reasonable alphas with up to a value of around ``100``. - (3) Conversion to ``float64`` is done during augmentation, hence ``uint64``, ``int64``, and ``float128`` support cannot be guaranteed. - (4) Does not make sense for contrast adjustments. Parameters ---------- arr : numpy.ndarray Array for which to adjust the contrast. Dtype ``uint8`` is fastest. alpha : number Multiplier to linearly pronounce (``>1.0``), dampen (``0.0`` to ``1.0``) or invert (``<0.0``) the difference between each pixel value and the dtype's center value, e.g. ``127`` for ``uint8``. Returns ------- numpy.ndarray Array with adjusted contrast. """ # pylint: disable=no-else-return if arr.size == 0: return np.copy(arr) # int8 is also possible according to docs # https://docs.opencv.org/3.0-beta/modules/core/doc/operations_on_arrays.html#cv2.LUT , # but here it seemed like `d` was 0 for CV_8S, causing that to fail if arr.dtype.name == "uint8": min_value, center_value, max_value = \ iadt.get_value_range_of_dtype(arr.dtype) # TODO get rid of this int(...) center_value = int(center_value) value_range = np.arange(0, 256, dtype=np.float32) # 127 + alpha*(I_ij-127) # using np.float32(.) here still works when the input is a numpy array # of size 1 alpha = np.float32(alpha) table = center_value + alpha * (value_range - center_value) table = np.clip(table, min_value, max_value).astype(arr.dtype) arr_aug = ia.apply_lut(arr, table) return arr_aug else: input_dtype = arr.dtype _min_value, center_value, _max_value = \ iadt.get_value_range_of_dtype(input_dtype) # TODO get rid of this int(...) if input_dtype.kind in ["u", "i"]: center_value = int(center_value) image_aug = (center_value + alpha * (arr.astype(np.float64)-center_value)) image_aug = iadt.restore_dtypes_(image_aug, input_dtype) return image_aug class GammaContrast(_ContrastFuncWrapper): """ Adjust image contrast by scaling pixel values to ``255*((v/255)**gamma)``. Values in the range ``gamma=(0.5, 2.0)`` seem to be sensible. **Supported dtypes**: See :func:`~imgaug.augmenters.contrast.adjust_contrast_gamma`. Parameters ---------- gamma : number or tuple of number or list of number or imgaug.parameters.StochasticParameter, optional Exponent for the contrast adjustment. Higher values darken the image. * If a number, then that value will be used for all images. * If a tuple ``(a, b)``, then a value from the range ``[a, b]`` will be used per image. * If a list, then a random value will be sampled from that list per image. * If a ``StochasticParameter``, then a value will be sampled per image from that parameter. per_channel : bool or float, optional Whether to use the same value for all channels (``False``) or to sample a new value for each channel (``True``). If this value is a float ``p``, then for ``p`` percent of all images `per_channel` will be treated as ``True``, otherwise as ``False``. seed : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. name : None or str, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. random_state : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional Old name for parameter `seed`. Its usage will not yet cause a deprecation warning, but it is still recommended to use `seed` now. Outdated since 0.4.0. deterministic : bool, optional Deprecated since 0.4.0. See method ``to_deterministic()`` for an alternative and for details about what the "deterministic mode" actually does. Examples -------- >>> import imgaug.augmenters as iaa >>> aug = iaa.GammaContrast((0.5, 2.0)) Modify the contrast of images according to ``255*((v/255)**gamma)``, where ``v`` is a pixel value and ``gamma`` is sampled uniformly from the interval ``[0.5, 2.0]`` (once per image). >>> aug = iaa.GammaContrast((0.5, 2.0), per_channel=True) Same as in the previous example, but ``gamma`` is sampled once per image *and* channel. """ def __init__(self, gamma=(0.7, 1.7), per_channel=False, seed=None, name=None, random_state="deprecated", deterministic="deprecated"): params1d = [iap.handle_continuous_param( gamma, "gamma", value_range=None, tuple_to_uniform=True, list_to_choice=True)] func = adjust_contrast_gamma super(GammaContrast, self).__init__( func, params1d, per_channel, dtypes_allowed=["uint8", "uint16", "uint32", "uint64", "int8", "int16", "int32", "int64", "float16", "float32", "float64"], dtypes_disallowed=["float96", "float128", "float256", "bool"], seed=seed, name=name, random_state=random_state, deterministic=deterministic) class SigmoidContrast(_ContrastFuncWrapper): """ Adjust image contrast to ``255*1/(1+exp(gain*(cutoff-I_ij/255)))``. Values in the range ``gain=(5, 20)`` and ``cutoff=(0.25, 0.75)`` seem to be sensible. A combination of ``gain=5.5`` and ``cutof=0.45`` is fairly close to the identity function. **Supported dtypes**: See :func:`~imgaug.augmenters.contrast.adjust_contrast_sigmoid`. Parameters ---------- gain : number or tuple of number or list of number or imgaug.parameters.StochasticParameter, optional Multiplier for the sigmoid function's output. Higher values lead to quicker changes from dark to light pixels. * If a number, then that value will be used for all images. * If a tuple ``(a, b)``, then a value from the interval ``[a, b]`` will be sampled uniformly per image. * If a list, then a random value will be sampled from that list per image. * If a ``StochasticParameter``, then a value will be sampled per image from that parameter. cutoff : number or tuple of number or list of number or imgaug.parameters.StochasticParameter, optional Cutoff that shifts the sigmoid function in horizontal direction. Higher values mean that the switch from dark to light pixels happens later, i.e. the pixels will remain darker. * If a number, then that value will be used for all images. * If a tuple ``(a, b)``, then a value from the range ``[a, b]`` will be used per image. * If a list, then a random value will be sampled from that list per image. * If a ``StochasticParameter``, then a value will be sampled per image from that parameter. per_channel : bool or float, optional Whether to use the same value for all channels (``False``) or to sample a new value for each channel (``True``). If this value is a float ``p``, then for ``p`` percent of all images `per_channel` will be treated as ``True``, otherwise as ``False``. seed : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. name : None or str, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. random_state : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional Old name for parameter `seed`. Its usage will not yet cause a deprecation warning, but it is still recommended to use `seed` now. Outdated since 0.4.0. deterministic : bool, optional Deprecated since 0.4.0. See method ``to_deterministic()`` for an alternative and for details about what the "deterministic mode" actually does. Examples -------- >>> import imgaug.augmenters as iaa >>> aug = iaa.SigmoidContrast(gain=(3, 10), cutoff=(0.4, 0.6)) Modify the contrast of images according to ``255*1/(1+exp(gain*(cutoff-v/255)))``, where ``v`` is a pixel value, ``gain`` is sampled uniformly from the interval ``[3, 10]`` (once per image) and ``cutoff`` is sampled uniformly from the interval ``[0.4, 0.6]`` (also once per image). >>> aug = iaa.SigmoidContrast( >>> gain=(3, 10), cutoff=(0.4, 0.6), per_channel=True) Same as in the previous example, but ``gain`` and ``cutoff`` are each sampled once per image *and* channel. """ def __init__(self, gain=(5, 6), cutoff=(0.3, 0.6), per_channel=False, seed=None, name=None, random_state="deprecated", deterministic="deprecated"): # TODO add inv parameter? params1d = [ iap.handle_continuous_param( gain, "gain", value_range=(0, None), tuple_to_uniform=True, list_to_choice=True), iap.handle_continuous_param( cutoff, "cutoff", value_range=(0, 1.0), tuple_to_uniform=True, list_to_choice=True) ] func = adjust_contrast_sigmoid super(SigmoidContrast, self).__init__( func, params1d, per_channel, dtypes_allowed=["uint8", "uint16", "uint32", "uint64", "int8", "int16", "int32", "int64", "float16", "float32", "float64"], dtypes_disallowed=["float96", "float128", "float256", "bool"], seed=seed, name=name, random_state=random_state, deterministic=deterministic) class LogContrast(_ContrastFuncWrapper): """Adjust image contrast by scaling pixels to ``255*gain*log_2(1+v/255)``. This augmenter is fairly similar to ``imgaug.augmenters.arithmetic.Multiply``. **Supported dtypes**: See :func:`~imgaug.augmenters.contrast.adjust_contrast_log`. Parameters ---------- gain : number or tuple of number or list of number or imgaug.parameters.StochasticParameter, optional Multiplier for the logarithm result. Values around ``1.0`` lead to a contrast-adjusted images. Values above ``1.0`` quickly lead to partially broken images due to exceeding the datatype's value range. * If a number, then that value will be used for all images. * If a tuple ``(a, b)``, then a value from the interval ``[a, b]`` will uniformly sampled be used per image. * If a list, then a random value will be sampled from that list per image. * If a ``StochasticParameter``, then a value will be sampled per image from that parameter. per_channel : bool or float, optional Whether to use the same value for all channels (``False``) or to sample a new value for each channel (``True``). If this value is a float ``p``, then for ``p`` percent of all images `per_channel` will be treated as ``True``, otherwise as ``False``. seed : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. name : None or str, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. random_state : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional Old name for parameter `seed`. Its usage will not yet cause a deprecation warning, but it is still recommended to use `seed` now. Outdated since 0.4.0. deterministic : bool, optional Deprecated since 0.4.0. See method ``to_deterministic()`` for an alternative and for details about what the "deterministic mode" actually does. Examples -------- >>> import imgaug.augmenters as iaa >>> aug = iaa.LogContrast(gain=(0.6, 1.4)) Modify the contrast of images according to ``255*gain*log_2(1+v/255)``, where ``v`` is a pixel value and ``gain`` is sampled uniformly from the interval ``[0.6, 1.4]`` (once per image). >>> aug = iaa.LogContrast(gain=(0.6, 1.4), per_channel=True) Same as in the previous example, but ``gain`` is sampled once per image *and* channel. """ def __init__(self, gain=(0.4, 1.6), per_channel=False, seed=None, name=None, random_state="deprecated", deterministic="deprecated"): # TODO add inv parameter? params1d = [iap.handle_continuous_param( gain, "gain", value_range=(0, None), tuple_to_uniform=True, list_to_choice=True)] func = adjust_contrast_log super(LogContrast, self).__init__( func, params1d, per_channel, dtypes_allowed=["uint8", "uint16", "uint32", "uint64", "int8", "int16", "int32", "int64", "float16", "float32", "float64"], dtypes_disallowed=["float96", "float128", "float256", "bool"], seed=seed, name=name, random_state=random_state, deterministic=deterministic) class LinearContrast(_ContrastFuncWrapper): """Adjust contrast by scaling each pixel to ``127 + alpha*(v-127)``. **Supported dtypes**: See :func:`~imgaug.augmenters.contrast.adjust_contrast_linear`. Parameters ---------- alpha : number or tuple of number or list of number or imgaug.parameters.StochasticParameter, optional Multiplier to linearly pronounce (``>1.0``), dampen (``0.0`` to ``1.0``) or invert (``<0.0``) the difference between each pixel value and the dtype's center value, e.g. ``127`` for ``uint8``. * If a number, then that value will be used for all images. * If a tuple ``(a, b)``, then a value from the interval ``[a, b]`` will be used per image. * If a list, then a random value will be sampled from that list per image. * If a ``StochasticParameter``, then a value will be sampled per image from that parameter. per_channel : bool or float, optional Whether to use the same value for all channels (``False``) or to sample a new value for each channel (``True``). If this value is a float ``p``, then for ``p`` percent of all images `per_channel` will be treated as ``True``, otherwise as ``False``. seed : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. name : None or str, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. random_state : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional Old name for parameter `seed`. Its usage will not yet cause a deprecation warning, but it is still recommended to use `seed` now. Outdated since 0.4.0. deterministic : bool, optional Deprecated since 0.4.0. See method ``to_deterministic()`` for an alternative and for details about what the "deterministic mode" actually does. Examples -------- >>> import imgaug.augmenters as iaa >>> aug = iaa.LinearContrast((0.4, 1.6)) Modify the contrast of images according to `127 + alpha*(v-127)``, where ``v`` is a pixel value and ``alpha`` is sampled uniformly from the interval ``[0.4, 1.6]`` (once per image). >>> aug = iaa.LinearContrast((0.4, 1.6), per_channel=True) Same as in the previous example, but ``alpha`` is sampled once per image *and* channel. """ def __init__(self, alpha=(0.6, 1.4), per_channel=False, seed=None, name=None, random_state="deprecated", deterministic="deprecated"): params1d = [ iap.handle_continuous_param( alpha, "alpha", value_range=None, tuple_to_uniform=True, list_to_choice=True) ] func = adjust_contrast_linear super(LinearContrast, self).__init__( func, params1d, per_channel, dtypes_allowed=["uint8", "uint16", "uint32", "int8", "int16", "int32", "float16", "float32", "float64"], dtypes_disallowed=["uint64", "int64", "float96", "float128", "float256", "bool"], seed=seed, name=name, random_state=random_state, deterministic=deterministic) # TODO maybe offer the other contrast augmenters also wrapped in this, similar # to CLAHE and HistogramEqualization? # this is essentially tested by tests for CLAHE class _IntensityChannelBasedApplier(object): RGB = color_lib.CSPACE_RGB BGR = color_lib.CSPACE_BGR HSV = color_lib.CSPACE_HSV HLS = color_lib.CSPACE_HLS Lab = color_lib.CSPACE_Lab _CHANNEL_MAPPING = { HSV: 2, HLS: 1, Lab: 0 } def __init__(self, from_colorspace, to_colorspace): super(_IntensityChannelBasedApplier, self).__init__() # TODO maybe add CIE, Luv? valid_from_colorspaces = [self.RGB, self.BGR, self.Lab, self.HLS, self.HSV] assert from_colorspace in valid_from_colorspaces, ( "Expected 'from_colorspace' to be one of %s, got %s." % ( valid_from_colorspaces, from_colorspace)) valid_to_colorspaces = [self.Lab, self.HLS, self.HSV] assert to_colorspace in valid_to_colorspaces, ( "Expected 'to_colorspace' to be one of %s, got %s." % ( valid_to_colorspaces, to_colorspace)) self.from_colorspace = from_colorspace self.to_colorspace = to_colorspace def apply(self, images, random_state, parents, hooks, func): input_was_array = ia.is_np_array(images) rss = random_state.duplicate(3) # normalize images # (H, W, 1) will be used directly in AllChannelsCLAHE # (H, W, 3) will be converted to target colorspace in the next # block # (H, W, 4) will be reduced to (H, W, 3) (remove 4th channel) and # converted to target colorspace in next block # (H, W, ) will raise a warning and be treated channelwise by # AllChannelsCLAHE images_normalized = [] images_change_cs = [] images_change_cs_indices = [] for i, image in enumerate(images): nb_channels = image.shape[2] if nb_channels == 1: images_normalized.append(image) elif nb_channels == 3: images_normalized.append(None) images_change_cs.append(image) images_change_cs_indices.append(i) elif nb_channels == 4: # assume that 4th channel is an alpha channel, e.g. in RGBA images_normalized.append(None) images_change_cs.append(image[..., 0:3]) images_change_cs_indices.append(i) else: ia.warn( "Got image with %d channels in " "_IntensityChannelBasedApplier (parents: %s), " "expected 0, 1, 3 or 4 channels." % ( nb_channels, ", ".join( parent.name for parent in parents))) images_normalized.append(image) # convert colorspaces of normalized 3-channel images images_after_color_conversion = [None] * len(images_normalized) if len(images_change_cs) > 0: images_new_cs = color_lib.change_colorspaces_( images_change_cs, to_colorspaces=self.to_colorspace, from_colorspaces=self.from_colorspace) for image_new_cs, target_idx in zip(images_new_cs, images_change_cs_indices): chan_idx = self._CHANNEL_MAPPING[self.to_colorspace] images_normalized[target_idx] = image_new_cs[ ..., chan_idx:chan_idx+1] images_after_color_conversion[target_idx] = image_new_cs # apply function channelwise images_aug = func(images_normalized, rss[1]) # denormalize result = [] images_change_cs = [] images_change_cs_indices = [] gen = enumerate(zip(images, images_after_color_conversion, images_aug)) for i, (image, image_conv, image_aug) in gen: nb_channels = image.shape[2] if nb_channels in [3, 4]: chan_idx = self._CHANNEL_MAPPING[self.to_colorspace] image_tmp = image_conv image_tmp[..., chan_idx:chan_idx+1] = image_aug result.append(None if nb_channels == 3 else image[..., 3:4]) images_change_cs.append(image_tmp) images_change_cs_indices.append(i) else: result.append(image_aug) # invert colorspace conversion if len(images_change_cs) > 0: images_new_cs = color_lib.change_colorspaces_( images_change_cs, to_colorspaces=self.from_colorspace, from_colorspaces=self.to_colorspace) for image_new_cs, target_idx in zip(images_new_cs, images_change_cs_indices): if result[target_idx] is None: result[target_idx] = image_new_cs else: # input image had four channels, 4th channel is already # in result result[target_idx] = np.dstack((image_new_cs, result[target_idx])) # convert to array if necessary if input_was_array: result = np.array(result, dtype=result[0].dtype) return result def get_parameters(self): """See :func:`~imgaug.augmenters.meta.Augmenter.get_parameters`.""" return [self.from_colorspace, self.to_colorspace] # TODO add parameter `tile_grid_size_percent` class AllChannelsCLAHE(meta.Augmenter): """Apply CLAHE to all channels of images in their original colorspaces. CLAHE (Contrast Limited Adaptive Histogram Equalization) performs histogram equilization within image patches, i.e. over local neighbourhoods. In contrast to ``imgaug.augmenters.contrast.CLAHE``, this augmenter operates directly on all channels of the input images. It does not perform any colorspace transformations and does not focus on specific channels (e.g. ``L`` in ``Lab`` colorspace). **Supported dtypes**: * ``uint8``: yes; fully tested * ``uint16``: yes; tested * ``uint32``: no (1) * ``uint64``: no (2) * ``int8``: no (2) * ``int16``: no (2) * ``int32``: no (2) * ``int64``: no (2) * ``float16``: no (2) * ``float32``: no (2) * ``float64``: no (2) * ``float128``: no (1) * ``bool``: no (1) - (1) rejected by cv2 - (2) results in error in cv2: ``cv2.error: OpenCV(3.4.2) (...)/clahe.cpp:351: error: (-215:Assertion failed) src.type() == (((0) & ((1 << 3) - 1)) + (((1)-1) << 3)) || _src.type() == (((2) & ((1 << 3) - 1)) + (((1)-1) << 3)) in function 'apply'`` Parameters ---------- clip_limit : number or tuple of number or list of number or imgaug.parameters.StochasticParameter, optional See ``imgaug.augmenters.contrast.CLAHE``. tile_grid_size_px : int or tuple of int or list of int or imgaug.parameters.StochasticParameter or tuple of tuple of int or tuple of list of int or tuple of imgaug.parameters.StochasticParameter, optional See ``imgaug.augmenters.contrast.CLAHE``. tile_grid_size_px_min : int, optional See ``imgaug.augmenters.contrast.CLAHE``. per_channel : bool or float, optional Whether to use the same value for all channels (``False``) or to sample a new value for each channel (``True``). If this value is a float ``p``, then for ``p`` percent of all images `per_channel` will be treated as ``True``, otherwise as ``False``. seed : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. name : None or str, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. random_state : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional Old name for parameter `seed`. Its usage will not yet cause a deprecation warning, but it is still recommended to use `seed` now. Outdated since 0.4.0. deterministic : bool, optional Deprecated since 0.4.0. See method ``to_deterministic()`` for an alternative and for details about what the "deterministic mode" actually does. Examples -------- >>> import imgaug.augmenters as iaa >>> aug = iaa.AllChannelsCLAHE() Create an augmenter that applies CLAHE to all channels of input images. >>> aug = iaa.AllChannelsCLAHE(clip_limit=(1, 10)) Same as in the previous example, but the `clip_limit` used by CLAHE is uniformly sampled per image from the interval ``[1, 10]``. Some images will therefore have stronger contrast than others (i.e. higher clip limit values). >>> aug = iaa.AllChannelsCLAHE(clip_limit=(1, 10), per_channel=True) Same as in the previous example, but the `clip_limit` is sampled per image *and* channel, leading to different levels of contrast for each channel. """ def __init__(self, clip_limit=(0.1, 8), tile_grid_size_px=(3, 12), tile_grid_size_px_min=3, per_channel=False, seed=None, name=None, random_state="deprecated", deterministic="deprecated"): super(AllChannelsCLAHE, self).__init__( seed=seed, name=name, random_state=random_state, deterministic=deterministic) self.clip_limit = iap.handle_continuous_param( clip_limit, "clip_limit", value_range=(0+1e-4, None), tuple_to_uniform=True, list_to_choice=True) self.tile_grid_size_px = iap.handle_discrete_kernel_size_param( tile_grid_size_px, "tile_grid_size_px", value_range=(0, None), allow_floats=False) self.tile_grid_size_px_min = tile_grid_size_px_min self.per_channel = iap.handle_probability_param(per_channel, "per_channel") # Added in 0.4.0. def _augment_batch_(self, batch, random_state, parents, hooks): if batch.images is None: return batch images = batch.images iadt.gate_dtypes( images, allowed=["uint8", "uint16"], disallowed=["bool", "uint32", "uint64", "uint128", "uint256", "int8", "int16", "int32", "int64", "int128", "int256", "float16", "float32", "float64", "float96", "float128", "float256"], augmenter=self) nb_images = len(images) nb_channels = meta.estimate_max_number_of_channels(images) mode = "single" if self.tile_grid_size_px[1] is None else "two" rss = random_state.duplicate(3 if mode == "single" else 4) per_channel = self.per_channel.draw_samples((nb_images,), random_state=rss[0]) clip_limit = self.clip_limit.draw_samples((nb_images, nb_channels), random_state=rss[1]) tile_grid_size_px_h = self.tile_grid_size_px[0].draw_samples( (nb_images, nb_channels), random_state=rss[2]) if mode == "single": tile_grid_size_px_w = tile_grid_size_px_h else: tile_grid_size_px_w = self.tile_grid_size_px[1].draw_samples( (nb_images, nb_channels), random_state=rss[3]) tile_grid_size_px_w = np.maximum(tile_grid_size_px_w, self.tile_grid_size_px_min) tile_grid_size_px_h = np.maximum(tile_grid_size_px_h, self.tile_grid_size_px_min) gen = enumerate(zip(images, clip_limit, tile_grid_size_px_h, tile_grid_size_px_w, per_channel)) for i, (image, clip_limit_i, tgs_px_h_i, tgs_px_w_i, pchannel_i) in gen: if image.size == 0: continue nb_channels = image.shape[2] c_param = 0 image_warped = [] for c in sm.xrange(nb_channels): if tgs_px_w_i[c_param] > 1 or tgs_px_h_i[c_param] > 1: clahe = cv2.createCLAHE( clipLimit=clip_limit_i[c_param], tileGridSize=(tgs_px_w_i[c_param], tgs_px_h_i[c_param]) ) channel_warped = clahe.apply( _normalize_cv2_input_arr_(image[..., c]) ) image_warped.append(channel_warped) else: image_warped.append(image[..., c]) if pchannel_i > 0.5: c_param += 1 # combine channels to one image image_warped = np.stack(image_warped, axis=-1) batch.images[i] = image_warped return batch def get_parameters(self): """See :func:`~imgaug.augmenters.meta.Augmenter.get_parameters`.""" return [self.clip_limit, self.tile_grid_size_px, self.tile_grid_size_px_min, self.per_channel] class CLAHE(meta.Augmenter): """Apply CLAHE to L/V/L channels in HLS/HSV/Lab colorspaces. This augmenter applies CLAHE (Contrast Limited Adaptive Histogram Equalization) to images, a form of histogram equalization that normalizes within local image patches. The augmenter transforms input images to a target colorspace (e.g. ``Lab``), extracts an intensity-related channel from the converted images (e.g. ``L`` for ``Lab``), applies CLAHE to the channel and then converts the resulting image back to the original colorspace. Grayscale images (images without channel axis or with only one channel axis) are automatically handled, `from_colorspace` does not have to be adjusted for them. For images with four channels (e.g. ``RGBA``), the fourth channel is ignored in the colorspace conversion (e.g. from an ``RGBA`` image, only the ``RGB`` part is converted, normalized, converted back and concatenated with the input ``A`` channel). Images with unusual channel numbers (2, 5 or more than 5) are normalized channel-by-channel (same behaviour as ``AllChannelsCLAHE``, though a warning will be raised). If you want to apply CLAHE to each channel of the original input image's colorspace (without any colorspace conversion), use ``imgaug.augmenters.contrast.AllChannelsCLAHE`` instead. **Supported dtypes**: * ``uint8``: yes; fully tested * ``uint16``: no (1) * ``uint32``: no (1) * ``uint64``: no (1) * ``int8``: no (1) * ``int16``: no (1) * ``int32``: no (1) * ``int64``: no (1) * ``float16``: no (1) * ``float32``: no (1) * ``float64``: no (1) * ``float128``: no (1) * ``bool``: no (1) - (1) This augmenter uses :class:`~imgaug.augmenters.color.ChangeColorspace`, which is currently limited to ``uint8``. Parameters ---------- clip_limit : number or tuple of number or list of number or imgaug.parameters.StochasticParameter, optional Clipping limit. Higher values result in stronger contrast. OpenCV uses a default of ``40``, though values around ``5`` seem to already produce decent contrast. * If a number, then that value will be used for all images. * If a tuple ``(a, b)``, then a value from the range ``[a, b]`` will be used per image. * If a list, then a random value will be sampled from that list per image. * If a ``StochasticParameter``, then a value will be sampled per image from that parameter. tile_grid_size_px : int or tuple of int or list of int or imgaug.parameters.StochasticParameter or tuple of tuple of int or tuple of list of int or tuple of imgaug.parameters.StochasticParameter, optional Kernel size, i.e. size of each local neighbourhood in pixels. * If an ``int``, then that value will be used for all images for both kernel height and width. * If a tuple ``(a, b)``, then a value from the discrete interval ``[a..b]`` will be uniformly sampled per image. * If a list, then a random value will be sampled from that list per image and used for both kernel height and width. * If a ``StochasticParameter``, then a value will be sampled per image from that parameter per image and used for both kernel height and width. * If a tuple of tuple of ``int`` given as ``((a, b), (c, d))``, then two values will be sampled independently from the discrete ranges ``[a..b]`` and ``[c..d]`` per image and used as the kernel height and width. * If a tuple of lists of ``int``, then two values will be sampled independently per image, one from the first list and one from the second, and used as the kernel height and width. * If a tuple of ``StochasticParameter``, then two values will be sampled indepdently per image, one from the first parameter and one from the second, and used as the kernel height and width. tile_grid_size_px_min : int, optional Minimum kernel size in px, per axis. If the sampling results in a value lower than this minimum, it will be clipped to this value. from_colorspace : {"RGB", "BGR", "HSV", "HLS", "Lab"}, optional Colorspace of the input images. If any input image has only one or zero channels, this setting will be ignored and it will be assumed that the input is grayscale. If a fourth channel is present in an input image, it will be removed before the colorspace conversion and later re-added. See also :func:`~imgaug.augmenters.color.change_colorspace_` for details. to_colorspace : {"Lab", "HLS", "HSV"}, optional Colorspace in which to perform CLAHE. For ``Lab``, CLAHE will only be applied to the first channel (``L``), for ``HLS`` to the second (``L``) and for ``HSV`` to the third (``V``). To apply CLAHE to all channels of an input image (without colorspace conversion), see ``imgaug.augmenters.contrast.AllChannelsCLAHE``. seed : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. name : None or str, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. random_state : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional Old name for parameter `seed`. Its usage will not yet cause a deprecation warning, but it is still recommended to use `seed` now. Outdated since 0.4.0. deterministic : bool, optional Deprecated since 0.4.0. See method ``to_deterministic()`` for an alternative and for details about what the "deterministic mode" actually does. Examples -------- >>> import imgaug.augmenters as iaa >>> aug = iaa.CLAHE() Create a standard CLAHE augmenter. >>> aug = iaa.CLAHE(clip_limit=(1, 10)) Create a CLAHE augmenter with a clip limit uniformly sampled from ``[1..10]``, where ``1`` is rather low contrast and ``10`` is rather high contrast. >>> aug = iaa.CLAHE(tile_grid_size_px=(3, 21)) Create a CLAHE augmenter with kernel sizes of ``SxS``, where ``S`` is uniformly sampled from ``[3..21]``. Sampling happens once per image. >>> aug = iaa.CLAHE( >>> tile_grid_size_px=iap.Discretize(iap.Normal(loc=7, scale=2)), >>> tile_grid_size_px_min=3) Create a CLAHE augmenter with kernel sizes of ``SxS``, where ``S`` is sampled from ``N(7, 2)``, but does not go below ``3``. >>> aug = iaa.CLAHE(tile_grid_size_px=((3, 21), [3, 5, 7])) Create a CLAHE augmenter with kernel sizes of ``HxW``, where ``H`` is uniformly sampled from ``[3..21]`` and ``W`` is randomly picked from the list ``[3, 5, 7]``. >>> aug = iaa.CLAHE( >>> from_colorspace=iaa.CSPACE_BGR, >>> to_colorspace=iaa.CSPACE_HSV) Create a CLAHE augmenter that converts images from BGR colorspace to HSV colorspace and then applies the local histogram equalization to the ``V`` channel of the images (before converting back to ``BGR``). Alternatively, ``Lab`` (default) or ``HLS`` can be used as the target colorspace. Grayscale images (no channels / one channel) are never converted and are instead directly normalized (i.e. `from_colorspace` does not have to be changed for them). """ RGB = color_lib.CSPACE_RGB BGR = color_lib.CSPACE_BGR HSV = color_lib.CSPACE_HSV HLS = color_lib.CSPACE_HLS Lab = color_lib.CSPACE_Lab def __init__(self, clip_limit=(0.1, 8), tile_grid_size_px=(3, 12), tile_grid_size_px_min=3, from_colorspace=color_lib.CSPACE_RGB, to_colorspace=color_lib.CSPACE_Lab, seed=None, name=None, random_state="deprecated", deterministic="deprecated"): super(CLAHE, self).__init__( seed=seed, name=name, random_state=random_state, deterministic=deterministic) self.all_channel_clahe = AllChannelsCLAHE( clip_limit=clip_limit, tile_grid_size_px=tile_grid_size_px, tile_grid_size_px_min=tile_grid_size_px_min, name="%s_AllChannelsCLAHE" % (name,)) self.intensity_channel_based_applier = _IntensityChannelBasedApplier( from_colorspace, to_colorspace) # Added in 0.4.0. def _augment_batch_(self, batch, random_state, parents, hooks): if batch.images is None: return batch images = batch.images iadt.gate_dtypes( images, allowed=["uint8"], disallowed=["bool", "uint16", "uint32", "uint64", "uint128", "uint256", "int8", "int16", "int32", "int64", "int128", "int256", "float16", "float32", "float64", "float96", "float128", "float256"], augmenter=self) def _augment_all_channels_clahe(images_normalized, random_state_derived): # pylint: disable=protected-access # TODO would .augment_batch() be sufficient here? batch_imgs = _BatchInAugmentation( images=images_normalized) return self.all_channel_clahe._augment_batch_( batch_imgs, random_state_derived, parents + [self], hooks ).images batch.images = self.intensity_channel_based_applier.apply( images, random_state, parents + [self], hooks, _augment_all_channels_clahe) return batch def get_parameters(self): """See :func:`~imgaug.augmenters.meta.Augmenter.get_parameters`.""" ac_clahe = self.all_channel_clahe intb_applier = self.intensity_channel_based_applier return [ ac_clahe.clip_limit, ac_clahe.tile_grid_size_px, ac_clahe.tile_grid_size_px_min ] + intb_applier.get_parameters() class AllChannelsHistogramEqualization(meta.Augmenter): """ Apply Histogram Eq. to all channels of images in their original colorspaces. In contrast to ``imgaug.augmenters.contrast.HistogramEqualization``, this augmenter operates directly on all channels of the input images. It does not perform any colorspace transformations and does not focus on specific channels (e.g. ``L`` in ``Lab`` colorspace). **Supported dtypes**: * ``uint8``: yes; fully tested * ``uint16``: no (1) * ``uint32``: no (2) * ``uint64``: no (1) * ``int8``: no (1) * ``int16``: no (1) * ``int32``: no (1) * ``int64``: no (1) * ``float16``: no (2) * ``float32``: no (1) * ``float64``: no (1) * ``float128``: no (2) * ``bool``: no (1) - (1) causes cv2 error: ``cv2.error: OpenCV(3.4.5) (...)/histogram.cpp:3345: error: (-215:Assertion failed) src.type() == CV_8UC1 in function 'equalizeHist'`` - (2) rejected by cv2 Parameters ---------- seed : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. name : None or str, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. random_state : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional Old name for parameter `seed`. Its usage will not yet cause a deprecation warning, but it is still recommended to use `seed` now. Outdated since 0.4.0. deterministic : bool, optional Deprecated since 0.4.0. See method ``to_deterministic()`` for an alternative and for details about what the "deterministic mode" actually does. Examples -------- >>> import imgaug.augmenters as iaa >>> aug = iaa.AllChannelsHistogramEqualization() Create an augmenter that applies histogram equalization to all channels of input images in the original colorspaces. >>> aug = iaa.Alpha((0.0, 1.0), iaa.AllChannelsHistogramEqualization()) Same as in the previous example, but alpha-blends the contrast-enhanced augmented images with the original input images using random blend strengths. This leads to random strengths of the contrast adjustment. """ def __init__(self, seed=None, name=None, random_state="deprecated", deterministic="deprecated"): super(AllChannelsHistogramEqualization, self).__init__( seed=seed, name=name, random_state=random_state, deterministic=deterministic) # Added in 0.4.0. def _augment_batch_(self, batch, random_state, parents, hooks): if batch.images is None: return batch images = batch.images iadt.gate_dtypes( images, allowed=["uint8"], disallowed=["bool", "uint16", "uint32", "uint64", "uint128", "uint256", "int8", "int16", "int32", "int64", "int128", "int256", "float16", "float32", "float64", "float96", "float128", "float256"], augmenter=self) for i, image in enumerate(images): if image.size == 0: continue image_warped = [ cv2.equalizeHist(_normalize_cv2_input_arr_(image[..., c])) for c in sm.xrange(image.shape[2])] image_warped = np.array(image_warped, dtype=image_warped[0].dtype) image_warped = image_warped.transpose((1, 2, 0)) batch.images[i] = image_warped return batch def get_parameters(self): """See :func:`~imgaug.augmenters.meta.Augmenter.get_parameters`.""" return [] class HistogramEqualization(meta.Augmenter): """ Apply Histogram Eq. to L/V/L channels of images in HLS/HSV/Lab colorspaces. This augmenter is similar to ``imgaug.augmenters.contrast.CLAHE``. The augmenter transforms input images to a target colorspace (e.g. ``Lab``), extracts an intensity-related channel from the converted images (e.g. ``L`` for ``Lab``), applies Histogram Equalization to the channel and then converts the resulting image back to the original colorspace. Grayscale images (images without channel axis or with only one channel axis) are automatically handled, `from_colorspace` does not have to be adjusted for them. For images with four channels (e.g. RGBA), the fourth channel is ignored in the colorspace conversion (e.g. from an ``RGBA`` image, only the ``RGB`` part is converted, normalized, converted back and concatenated with the input ``A`` channel). Images with unusual channel numbers (2, 5 or more than 5) are normalized channel-by-channel (same behaviour as ``AllChannelsHistogramEqualization``, though a warning will be raised). If you want to apply HistogramEqualization to each channel of the original input image's colorspace (without any colorspace conversion), use ``imgaug.augmenters.contrast.AllChannelsHistogramEqualization`` instead. **Supported dtypes**: * ``uint8``: yes; fully tested * ``uint16``: no (1) * ``uint32``: no (1) * ``uint64``: no (1) * ``int8``: no (1) * ``int16``: no (1) * ``int32``: no (1) * ``int64``: no (1) * ``float16``: no (1) * ``float32``: no (1) * ``float64``: no (1) * ``float128``: no (1) * ``bool``: no (1) - (1) This augmenter uses :class:`AllChannelsHistogramEqualization`, which only supports ``uint8``. Parameters ---------- from_colorspace : {"RGB", "BGR", "HSV", "HLS", "Lab"}, optional Colorspace of the input images. If any input image has only one or zero channels, this setting will be ignored and it will be assumed that the input is grayscale. If a fourth channel is present in an input image, it will be removed before the colorspace conversion and later re-added. See also :func:`~imgaug.augmenters.color.change_colorspace_` for details. to_colorspace : {"Lab", "HLS", "HSV"}, optional Colorspace in which to perform Histogram Equalization. For ``Lab``, the equalization will only be applied to the first channel (``L``), for ``HLS`` to the second (``L``) and for ``HSV`` to the third (``V``). To apply histogram equalization to all channels of an input image (without colorspace conversion), see ``imgaug.augmenters.contrast.AllChannelsHistogramEqualization``. seed : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. name : None or str, optional See :func:`~imgaug.augmenters.meta.Augmenter.__init__`. random_state : None or int or imgaug.random.RNG or numpy.random.Generator or numpy.random.BitGenerator or numpy.random.SeedSequence or numpy.random.RandomState, optional Old name for parameter `seed`. Its usage will not yet cause a deprecation warning, but it is still recommended to use `seed` now. Outdated since 0.4.0. deterministic : bool, optional Deprecated since 0.4.0. See method ``to_deterministic()`` for an alternative and for details about what the "deterministic mode" actually does. Examples -------- >>> import imgaug.augmenters as iaa >>> aug = iaa.HistogramEqualization() Create an augmenter that converts images to ``HLS``/``HSV``/``Lab`` colorspaces, extracts intensity-related channels (i.e. ``L``/``V``/``L``), applies histogram equalization to these channels and converts back to the input colorspace. >>> aug = iaa.Alpha((0.0, 1.0), iaa.HistogramEqualization()) Same as in the previous example, but alpha blends the result, leading to various strengths of contrast normalization. >>> aug = iaa.HistogramEqualization( >>> from_colorspace=iaa.CSPACE_BGR, >>> to_colorspace=iaa.CSPACE_HSV) Same as in the first example, but the colorspace of input images has to be ``BGR`` (instead of default ``RGB``) and the histogram equalization is applied to the ``V`` channel in ``HSV`` colorspace. """ RGB = color_lib.CSPACE_RGB BGR = color_lib.CSPACE_BGR HSV = color_lib.CSPACE_HSV HLS = color_lib.CSPACE_HLS Lab = color_lib.CSPACE_Lab def __init__(self, from_colorspace=color_lib.CSPACE_RGB, to_colorspace=color_lib.CSPACE_Lab, seed=None, name=None, random_state="deprecated", deterministic="deprecated"): super(HistogramEqualization, self).__init__( seed=seed, name=name, random_state=random_state, deterministic=deterministic) self.all_channel_histogram_equalization = \ AllChannelsHistogramEqualization( name="%s_AllChannelsHistogramEqualization" % (name,)) self.intensity_channel_based_applier = _IntensityChannelBasedApplier( from_colorspace, to_colorspace) # Added in 0.4.0. def _augment_batch_(self, batch, random_state, parents, hooks): if batch.images is None: return batch images = batch.images iadt.gate_dtypes( images, allowed=["uint8"], disallowed=["bool", "uint16", "uint32", "uint64", "uint128", "uint256", "int8", "int16", "int32", "int64", "int128", "int256", "float16", "float32", "float64", "float96", "float128", "float256"], augmenter=self) def _augment_all_channels_histogram_equalization(images_normalized, random_state_derived): # pylint: disable=protected-access # TODO would .augment_batch() be sufficient here batch_imgs = _BatchInAugmentation( images=images_normalized) return self.all_channel_histogram_equalization._augment_batch_( batch_imgs, random_state_derived, parents + [self], hooks ).images batch.images = self.intensity_channel_based_applier.apply( images, random_state, parents + [self], hooks, _augment_all_channels_histogram_equalization) return batch def get_parameters(self): """See :func:`~imgaug.augmenters.meta.Augmenter.get_parameters`.""" icb_applier = self.intensity_channel_based_applier return icb_applier.get_parameters()