Refactor augmenters to use new LUT functions

This patch switches all augmenters from cv2.lut() to
imgaug's apply_lut() and apply_lut_() functions.
This decreases code duplication and likely fixes
some bugs that were not yet discovered.

changelogs/master/added/20191230_standardized_lut.md

@@ -2,3 +2,6 @@
 
 * Added `imgaug.imgaug.apply_lut()`, which applies a lookup table to an image.
 * Added `imgaug.imgaug.apply_lut_()`. In-place version of `apply_lut()`.
+* Refactored all augmenters to use these new LUT functions.
+  This likely fixed some so-far undiscovered bugs in augmenters using LUT
+  tables.

imgaug/augmenters/arithmetic.py

@@ -153,23 +153,13 @@ def _add_scalar_to_uint8(image, value):
         result = []
         # TODO check if tile() is here actually needed
         tables = np.tile(
-            value_range[np.newaxis, :],
-            (nb_channels, 1)
-        ) + value[:, np.newaxis]
-        tables = np.clip(tables, 0, 255).astype(image.dtype)
-
-        for c, table in enumerate(tables):
-            result.append(cv2.LUT(image[..., c], table))
-
-        return np.stack(result, axis=-1)
+            value_range[:, np.newaxis],
+            (1, nb_channels)
+        ) + value[np.newaxis, :]
     else:
-        table = value_range + value
-        image_aug = cv2.LUT(
-            image,
-            iadt.clip_(table, 0, 255).astype(image.dtype))
-        if image_aug.ndim == 2 and image.ndim == 3:
-            image_aug = image_aug[..., np.newaxis]
-        return image_aug
+        tables = value_range + value
+    tables = np.clip(tables, 0, 255).astype(image.dtype)
+    return ia.apply_lut(image, tables)
 
 
 def _add_scalar_to_non_uint8(image, value):
@@ -443,23 +433,13 @@ def _multiply_scalar_to_uint8(image, multiplier):
         result = []
         # TODO check if tile() is here actually needed
         tables = np.tile(
-            value_range[np.newaxis, :],
-            (nb_channels, 1)
-        ) * multiplier[:, np.newaxis]
-        tables = np.clip(tables, 0, 255).astype(image.dtype)
-
-        for c, table in enumerate(tables):
-            arr_aug = cv2.LUT(image[..., c], table)
-            result.append(arr_aug)
-
-        return np.stack(result, axis=-1)
+            value_range[:, np.newaxis],
+            (1, nb_channels)
+        ) * multiplier[np.newaxis, :]
     else:
-        table = value_range * multiplier
-        image_aug = cv2.LUT(
-            image, np.clip(table, 0, 255).astype(image.dtype))
-        if image_aug.ndim == 2 and image.ndim == 3:
-            image_aug = image_aug[..., np.newaxis]
-        return image_aug
+        tables = value_range * multiplier
+    tables = np.clip(tables, 0, 255).astype(image.dtype)
+    return ia.apply_lut(image, tables)
 
 
 def _multiply_scalar_to_non_uint8(image, multiplier):
@@ -1198,17 +1178,9 @@ def _invert_bool(arr, min_value, max_value):
 
 def _invert_uint8_(arr, min_value, max_value, threshold,
                    invert_above_threshold):
-    if 0 in arr.shape:
-        return np.copy(arr)
-
-    if arr.flags["OWNDATA"] is False:
-        arr = np.copy(arr)
-    if arr.flags["C_CONTIGUOUS"] is False:
-        arr = np.ascontiguousarray(arr)
-
     table = _generate_table_for_invert_uint8(
         min_value, max_value, threshold, invert_above_threshold)
-    arr = cv2.LUT(arr, table, dst=arr)
+    arr = ia.apply_lut_(arr, table)
     return arr
 
 

imgaug/augmenters/blur.py

@@ -344,6 +344,7 @@ def blur_mean_shift_(image, spatial_window_radius, color_window_radius):
         image = np.tile(image, (1, 1, 3))
 
     # prevent image from becoming cv2.UMat
+    # TODO merge this with apply_lut() normalization/validation
     if image.flags["C_CONTIGUOUS"] is False:
         image = np.ascontiguousarray(image)
 

imgaug/augmenters/color.py

@@ -241,6 +241,7 @@ def _get_dst(image, from_to_cspace):
             # images that are views (e.g. image[..., 0:3]) and returns a
             # cv2.UMat instance instead of an array. So we check here first
             # if the array looks like it is non-contiguous or a view.
+            # TODO merge this with apply_lut() normalization/validation
             if image.flags["C_CONTIGUOUS"]:
                 return image
         return None
@@ -2315,11 +2316,13 @@ def _transform_image_cv2(cls, image_hsv, hue, saturation):
         # code with using cache (at best maybe 10% faster for 64x64):
         table_hue = cls._LUT_CACHE[0]
         table_saturation = cls._LUT_CACHE[1]
+        tables = [
+            table_hue[255+int(hue)],
+            table_saturation[255+int(saturation)]
+        ]
 
-        image_hsv[..., 0] = cv2.LUT(
-            image_hsv[..., 0], table_hue[255+int(hue)])
-        image_hsv[..., 1] = cv2.LUT(
-            image_hsv[..., 1], table_saturation[255+int(saturation)])
+        image_hsv[..., [0, 1]] = ia.apply_lut(image_hsv[..., [0, 1]],
+                                              tables)
 
         return image_hsv
 
@@ -3084,7 +3087,7 @@ def _generate_pixelwise_alpha_mask(cls, image_hsv, hue_to_alpha):
         hue = image_hsv[:, :, 0]
         table = hue_to_alpha * 255
         table = np.clip(np.round(table), 0, 255).astype(np.uint8)
-        mask = cv2.LUT(hue, table)
+        mask = ia.apply_lut(hue, table)
         return mask.astype(np.float32) / 255.0
 
     def get_parameters(self):
@@ -4055,22 +4058,20 @@ def quantize_uniform_(arr, nb_bins, to_bin_centers=True):
     if nb_bins == 256 or 0 in arr.shape:
         return arr
 
+    # TODO remove dtype check here? apply_lut_() does that already
     assert arr.dtype.name == "uint8", "Expected uint8 image, got %s." % (
         arr.dtype.name,)
     assert 2 <= nb_bins <= 256, (
         "Expected nb_bins to be in the discrete interval [2..256]. "
         "Got a value of %d instead." % (nb_bins,))
 
-    if arr.flags["C_CONTIGUOUS"] is False:
-        arr = np.ascontiguousarray(arr)
-
     table_class = (_QuantizeUniformCenterizedLUTTableSingleton
                    if to_bin_centers
                    else _QuantizeUniformNotCenterizedLUTTableSingleton)
     table = (table_class
              .get_instance()
              .get_for_nb_bins(nb_bins))
-    arr = cv2.LUT(arr, table, dst=arr)
+    arr = ia.apply_lut_(arr, table)
     return arr
 
 

imgaug/augmenters/contrast.py

@@ -146,8 +146,8 @@ def adjust_contrast_gamma(arr, gamma):
         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
+    # 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)
@@ -162,10 +162,8 @@ def adjust_contrast_gamma(arr, gamma):
         table = (min_value
                  + (value_range ** np.float32(gamma))
                  * dynamic_range)
-        arr_aug = cv2.LUT(
-            arr, np.clip(table, min_value, max_value).astype(arr.dtype))
-        if arr.ndim == 3 and arr_aug.ndim == 2:
-            return arr_aug[..., np.newaxis]
+        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)
 
@@ -232,8 +230,8 @@ def adjust_contrast_sigmoid(arr, gain, cutoff):
         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
+    # 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)
@@ -250,10 +248,8 @@ def adjust_contrast_sigmoid(arr, gain, cutoff):
         table = (min_value
                  + dynamic_range
                  * 1/(1 + np.exp(gain * (cutoff - value_range))))
-        arr_aug = cv2.LUT(
-            arr, np.clip(table, min_value, max_value).astype(arr.dtype))
-        if arr.ndim == 3 and arr_aug.ndim == 2:
-            return arr_aug[..., np.newaxis]
+        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)
 
@@ -319,8 +315,8 @@ def adjust_contrast_log(arr, gain):
         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
+    # 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)
@@ -334,10 +330,8 @@ def adjust_contrast_log(arr, gain):
         # of size 1
         gain = np.float32(gain)
         table = min_value + dynamic_range * gain * np.log2(1 + value_range)
-        arr_aug = cv2.LUT(
-            arr, np.clip(table, min_value, max_value).astype(arr.dtype))
-        if arr.ndim == 3 and arr_aug.ndim == 2:
-            return arr_aug[..., np.newaxis]
+        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)
 
@@ -391,8 +385,8 @@ def adjust_contrast_linear(arr, alpha):
         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
+    # 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)
@@ -406,10 +400,8 @@ def adjust_contrast_linear(arr, alpha):
         # of size 1
         alpha = np.float32(alpha)
         table = center_value + alpha * (value_range - center_value)
-        arr_aug = cv2.LUT(
-            arr, np.clip(table, min_value, max_value).astype(arr.dtype))
-        if arr.ndim == 3 and arr_aug.ndim == 2:
-            return arr_aug[..., np.newaxis]
+        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

imgaug/augmenters/pillike.py

@@ -282,13 +282,8 @@ def equalize_(image, mask=None):
 # note that this is supposed to be a non-PIL reimplementation of PIL's
 # equalize, which produces slightly different results from cv2.equalizeHist()
 def _equalize_no_pil_(image, mask=None):
-    flags = image.flags
-    if not flags["OWNDATA"]:
-        image = np.copy(image)
-    if not flags["C_CONTIGUOUS"]:
-        image = np.ascontiguousarray(image)
-
     nb_channels = 1 if image.ndim == 2 else image.shape[-1]
+    # TODO remove the first axis, no longer needed
     lut = np.empty((1, 256, nb_channels), dtype=np.int32)
 
     for c_idx in range(nb_channels):
@@ -312,9 +307,7 @@ def _equalize_no_pil_(image, mask=None):
         lut[0, 1:, c_idx] = n + cumsum[0:-1]
         lut[0, :, c_idx] //= int(step)
     lut = np.clip(lut, None, 255, out=lut).astype(np.uint8)
-    image = cv2.LUT(image, lut, dst=image)
-    if image.ndim == 2 and image.ndim == 3:
-        return image[..., np.newaxis]
+    image = ia.apply_lut_(image, lut)
     return image
 
 
@@ -420,7 +413,7 @@ def _autocontrast_no_pil(image, cutoff, ignore):  # noqa: C901
         # using [0] instead of [int(c_idx)] allows this to work with >4
         # channels
         if image.ndim == 2:
-            image_c = image
+            image_c = image[:, :, np.newaxis]
         else:
             image_c = image[:, :, c_idx:c_idx+1]
         h = cv2.calcHist([image_c], [0], None, [256], [0, 256])
@@ -488,7 +481,9 @@ def _autocontrast_no_pil(image, cutoff, ignore):  # noqa: C901
         #     ix = np.clip(ix, 0, 255).astype(np.uint8)
         #     lut = ix
 
-        result[:, :, c_idx] = cv2.LUT(image_c, lut)
+        # TODO change to a single call instead of one per channel
+        image_c_aug = ia.apply_lut(image_c, lut)
+        result[:, :, c_idx:c_idx+1] = image_c_aug
     if image.ndim == 2:
         return result[..., 0]
     return result