Merge pull request #1490 from girder/downsample

Zarr Tile Sink: Generate Downsampled Levels

large_image/tilesource/resample.py

@@ -0,0 +1,129 @@
+from enum import Enum
+from typing import Dict
+
+import numpy as np
+from PIL import Image
+
+
+class ResampleMethod(Enum):
+    PIL_NEAREST = Image.Resampling.NEAREST  # 0
+    PIL_LANCZOS = Image.Resampling.LANCZOS  # 1
+    PIL_BILINEAR = Image.Resampling.BILINEAR  # 2
+    PIL_BICUBIC = Image.Resampling.BICUBIC  # 3
+    PIL_BOX = Image.Resampling.BOX  # 4
+    PIL_HAMMING = Image.Resampling.HAMMING  # 5
+    PIL_MAX_ENUM = 5
+    NP_MEAN = 6
+    NP_MEDIAN = 7
+    NP_MODE = 8
+    NP_MAX = 9
+    NP_MIN = 10
+    NP_NEAREST = 11
+    NP_MAX_COLOR = 12
+    NP_MIN_COLOR = 13
+
+
+def pilResize(
+    tile: np.ndarray,
+    new_shape: Dict,
+    resample_method: ResampleMethod,
+) -> np.ndarray:
+    # Only NEAREST works for 16 bit images
+    img = Image.fromarray(tile)
+    resized_img = img.resize(
+        (new_shape['width'], new_shape['height']),
+        resample=resample_method.value,
+    )
+    result = np.array(resized_img).astype(tile.dtype)
+    return result
+
+
+def numpyResize(
+    tile: np.ndarray,
+    new_shape: Dict,
+    resample_method: ResampleMethod,
+) -> np.ndarray:
+    if resample_method == ResampleMethod.NP_NEAREST:
+        return tile[::2, ::2]
+    else:
+        if tile.shape[0] % 2 != 0:
+            tile = np.append(tile, np.expand_dims(tile[-1], axis=0), axis=0)
+        if tile.shape[1] % 2 != 0:
+            tile = np.append(tile, np.expand_dims(tile[:, -1], axis=1), axis=1)
+
+        pixel_selection = None
+        subarrays = np.asarray(
+            [
+                tile[0::2, 0::2],
+                tile[1::2, 0::2],
+                tile[0::2, 1::2],
+                tile[1::2, 1::2],
+            ],
+        )
+
+        if resample_method == ResampleMethod.NP_MEAN:
+            return np.mean(subarrays, axis=0).astype(tile.dtype)
+        elif resample_method == ResampleMethod.NP_MEDIAN:
+            return np.median(subarrays, axis=0).astype(tile.dtype)
+        elif resample_method == ResampleMethod.NP_MAX:
+            return np.max(subarrays, axis=0).astype(tile.dtype)
+        elif resample_method == ResampleMethod.NP_MIN:
+            return np.min(subarrays, axis=0).astype(tile.dtype)
+        elif resample_method == ResampleMethod.NP_MAX_COLOR:
+            summed = np.sum(subarrays, axis=3)
+            pixel_selection = np.argmax(summed, axis=0)
+        elif resample_method == ResampleMethod.NP_MIN_COLOR:
+            summed = np.sum(subarrays, axis=3)
+            pixel_selection = np.argmin(summed, axis=0)
+        elif resample_method == ResampleMethod.NP_MODE:
+            # if a pixel occurs twice in a set of four, it is a mode
+            # if no mode, default to pixel 0. check for minimal matches 1=2, 1=3, 2=3
+            pixel_selection = np.where(
+                (
+                    (subarrays[1] == subarrays[2]).all(axis=2) |
+                    (subarrays[1] == subarrays[3]).all(axis=2)
+                ),
+                1, np.where(
+                    (subarrays[2] == subarrays[3]).all(axis=2),
+                    2, 0,
+                ),
+            )
+
+        if pixel_selection is not None:
+            if len(tile.shape) > 2:
+                pixel_selection = np.expand_dims(pixel_selection, axis=2)
+                pixel_selection = np.repeat(pixel_selection, tile.shape[2], axis=2)
+            return np.choose(pixel_selection, subarrays).astype(tile.dtype)
+        else:
+            msg = f'Unknown resample method {resample_method}.'
+            raise ValueError(msg)
+
+
+def downsampleTileHalfRes(
+    tile: np.ndarray,
+    resample_method: ResampleMethod,
+) -> np.ndarray:
+    new_shape = {
+        'height': (tile.shape[0] + 1) // 2,
+        'width': (tile.shape[1] + 1) // 2,
+        'bands': 1,
+    }
+    if len(tile.shape) > 2:
+        new_shape['bands'] = tile.shape[-1]
+    if resample_method.value <= ResampleMethod.PIL_MAX_ENUM.value:
+        if new_shape['bands'] > 4:
+            result = np.empty(
+                (new_shape['height'], new_shape['width'], new_shape['bands']),
+                dtype=tile.dtype,
+            )
+            for band_index in range(new_shape['bands']):
+                result[(..., band_index)] = pilResize(
+                    tile[(..., band_index)],
+                    new_shape,
+                    resample_method,
+                )
+            return result
+        else:
+            return pilResize(tile, new_shape, resample_method)
+    else:
+        return numpyResize(tile, new_shape, resample_method)