Merge 9e415c7876 into 640f55a655

Tests/test_arro3.py

@@ -0,0 +1,277 @@
+from __future__ import annotations
+
+import json
+from typing import Any, NamedTuple
+
+import pytest
+
+from PIL import Image
+
+from .helper import (
+    assert_deep_equal,
+    assert_image_equal,
+    hopper,
+    is_big_endian,
+)
+
+TYPE_CHECKING = False
+if TYPE_CHECKING:
+    from arro3 import compute
+    from arro3.core import Array, DataType, Field, fixed_size_list_array
+else:
+    arro3 = pytest.importorskip("arro3", reason="Arro3 not installed")
+    from arro3 import compute
+    from arro3.core import Array, DataType, Field, fixed_size_list_array
+
+TEST_IMAGE_SIZE = (10, 10)
+
+
+def _test_img_equals_pyarray(
+    img: Image.Image, arr: Any, mask: list[int] | None, elts_per_pixel: int = 1
+) -> None:
+    assert img.height * img.width * elts_per_pixel == len(arr)
+    px = img.load()
+    assert px is not None
+    if elts_per_pixel > 1 and mask is None:
+        # have to do element-wise comparison when we're comparing
+        # flattened r,g,b,a to a pixel.
+        mask = list(range(elts_per_pixel))
+    for x in range(0, img.size[0], int(img.size[0] / 10)):
+        for y in range(0, img.size[1], int(img.size[1] / 10)):
+            if mask:
+                pixel = px[x, y]
+                assert isinstance(pixel, tuple)
+                for ix, elt in enumerate(mask):
+                    if elts_per_pixel == 1:
+                        assert pixel[ix] == arr[y * img.width + x].as_py()[elt]
+                    else:
+                        assert (
+                            pixel[ix]
+                            == arr[(y * img.width + x) * elts_per_pixel + elt].as_py()
+                        )
+            else:
+                assert_deep_equal(px[x, y], arr[y * img.width + x].as_py())
+
+
+def _test_img_equals_int32_pyarray(
+    img: Image.Image, arr: Any, mask: list[int] | None, elts_per_pixel: int = 1
+) -> None:
+    assert img.height * img.width * elts_per_pixel == len(arr)
+    px = img.load()
+    assert px is not None
+    if mask is None:
+        # have to do element-wise comparison when we're comparing
+        # flattened rgba in an uint32 to a pixel.
+        mask = list(range(elts_per_pixel))
+    for x in range(0, img.size[0], int(img.size[0] / 10)):
+        for y in range(0, img.size[1], int(img.size[1] / 10)):
+            pixel = px[x, y]
+            assert isinstance(pixel, tuple)
+            arr_pixel_int = arr[y * img.width + x].as_py()
+            arr_pixel_tuple = (
+                arr_pixel_int % 256,
+                (arr_pixel_int // 256) % 256,
+                (arr_pixel_int // 256**2) % 256,
+                (arr_pixel_int // 256**3),
+            )
+            if is_big_endian():
+                arr_pixel_tuple = arr_pixel_tuple[::-1]
+
+            for ix, elt in enumerate(mask):
+                assert pixel[ix] == arr_pixel_tuple[elt]
+
+
+fl_uint8_4_type = DataType.list(Field("_", DataType.uint8()).with_nullable(False), 4)
+
+
+@pytest.mark.parametrize(
+    "mode, dtype, mask",
+    (
+        ("L", DataType.uint8(), None),
+        ("I", DataType.int32(), None),
+        ("F", DataType.float32(), None),
+        ("LA", fl_uint8_4_type, [0, 3]),
+        ("RGB", fl_uint8_4_type, [0, 1, 2]),
+        ("RGBA", fl_uint8_4_type, None),
+        ("RGBX", fl_uint8_4_type, None),
+        ("CMYK", fl_uint8_4_type, None),
+        ("YCbCr", fl_uint8_4_type, [0, 1, 2]),
+        ("HSV", fl_uint8_4_type, [0, 1, 2]),
+    ),
+)
+def test_to_array(mode: str, dtype: DataType, mask: list[int] | None) -> None:
+    img = hopper(mode)
+
+    # Resize to non-square
+    img = img.crop((3, 0, 124, 127))
+    assert img.size == (121, 127)
+
+    arr = Array(img)  # type: ignore[call-overload]
+    _test_img_equals_pyarray(img, arr, mask)
+    assert arr.type == dtype
+
+    reloaded = Image.fromarrow(arr, mode, img.size)
+
+    assert reloaded
+
+    assert_image_equal(img, reloaded)
+
+
+def test_lifetime() -> None:
+    # valgrind shouldn't error out here.
+    # arrays should be accessible after the image is deleted.
+
+    img = hopper("L")
+
+    arr_1 = Array(img)  # type: ignore[call-overload]
+    arr_2 = Array(img)  # type: ignore[call-overload]
+
+    del img
+
+    assert compute.sum(arr_1).as_py() > 0
+    del arr_1
+
+    assert compute.sum(arr_2).as_py() > 0
+    del arr_2
+
+
+def test_lifetime2() -> None:
+    # valgrind shouldn't error out here.
+    # img should remain after the arrays are collected.
+
+    img = hopper("L")
+
+    arr_1 = Array(img)  # type: ignore[call-overload]
+    arr_2 = Array(img)  # type: ignore[call-overload]
+
+    assert compute.sum(arr_1).as_py() > 0
+    del arr_1
+
+    assert compute.sum(arr_2).as_py() > 0
+    del arr_2
+
+    img2 = img.copy()
+    px = img2.load()
+    assert px  # make mypy happy
+    assert isinstance(px[0, 0], int)
+
+
+class DataShape(NamedTuple):
+    dtype: DataType
+    # Strictly speaking, elt should be a pixel or pixel component, so
+    # list[uint8][4], float, int, uint32, uint8, etc.  But more
+    # correctly, it should be exactly the dtype from the line above.
+    elt: Any
+    elts_per_pixel: int
+
+
+UINT_ARR = DataShape(
+    dtype=fl_uint8_4_type,
+    elt=[1, 2, 3, 4],  # array of 4 uint8 per pixel
+    elts_per_pixel=1,  # only one array per pixel
+)
+
+UINT = DataShape(
+    dtype=DataType.uint8(),
+    elt=3,  # one uint8,
+    elts_per_pixel=4,  # but repeated 4x per pixel
+)
+
+UINT32 = DataShape(
+    dtype=DataType.uint32(),
+    elt=0xABCDEF45,  # one packed int, doesn't fit in a int32 > 0x80000000
+    elts_per_pixel=1,  # one per pixel
+)
+
+INT32 = DataShape(
+    dtype=DataType.uint32(),
+    elt=0x12CDEF45,  # one packed int
+    elts_per_pixel=1,  # one per pixel
+)
+
+
+@pytest.mark.parametrize(
+    "mode, data_tp, mask",
+    (
+        ("L", DataShape(DataType.uint8(), 3, 1), None),
+        ("I", DataShape(DataType.int32(), 1 << 24, 1), None),
+        ("F", DataShape(DataType.float32(), 3.14159, 1), None),
+        ("LA", UINT_ARR, [0, 3]),
+        ("LA", UINT, [0, 3]),
+        ("RGB", UINT_ARR, [0, 1, 2]),
+        ("RGBA", UINT_ARR, None),
+        ("CMYK", UINT_ARR, None),
+        ("YCbCr", UINT_ARR, [0, 1, 2]),
+        ("HSV", UINT_ARR, [0, 1, 2]),
+        ("RGB", UINT, [0, 1, 2]),
+        ("RGBA", UINT, None),
+        ("CMYK", UINT, None),
+        ("YCbCr", UINT, [0, 1, 2]),
+        ("HSV", UINT, [0, 1, 2]),
+    ),
+)
+def test_fromarray(mode: str, data_tp: DataShape, mask: list[int] | None) -> None:
+    (dtype, elt, elts_per_pixel) = data_tp
+
+    ct_pixels = TEST_IMAGE_SIZE[0] * TEST_IMAGE_SIZE[1]
+    if dtype == fl_uint8_4_type:
+        tmp_arr = Array(elt * (ct_pixels * elts_per_pixel), type=DataType.uint8())
+        arr = fixed_size_list_array(tmp_arr, 4)
+    else:
+        arr = Array([elt] * (ct_pixels * elts_per_pixel), type=dtype)
+    img = Image.fromarrow(arr, mode, TEST_IMAGE_SIZE)
+
+    _test_img_equals_pyarray(img, arr, mask, elts_per_pixel)
+
+
+@pytest.mark.parametrize(
+    "mode, data_tp, mask",
+    (
+        ("LA", UINT32, [0, 3]),
+        ("RGB", UINT32, [0, 1, 2]),
+        ("RGBA", UINT32, None),
+        ("CMYK", UINT32, None),
+        ("YCbCr", UINT32, [0, 1, 2]),
+        ("HSV", UINT32, [0, 1, 2]),
+        ("LA", INT32, [0, 3]),
+        ("RGB", INT32, [0, 1, 2]),
+        ("RGBA", INT32, None),
+        ("CMYK", INT32, None),
+        ("YCbCr", INT32, [0, 1, 2]),
+        ("HSV", INT32, [0, 1, 2]),
+    ),
+)
+def test_from_int32array(mode: str, data_tp: DataShape, mask: list[int] | None) -> None:
+    (dtype, elt, elts_per_pixel) = data_tp
+
+    ct_pixels = TEST_IMAGE_SIZE[0] * TEST_IMAGE_SIZE[1]
+    arr = Array([elt] * (ct_pixels * elts_per_pixel), type=dtype)
+    img = Image.fromarrow(arr, mode, TEST_IMAGE_SIZE)
+
+    _test_img_equals_int32_pyarray(img, arr, mask, elts_per_pixel)
+
+
+@pytest.mark.parametrize(
+    "mode, metadata",
+    (
+        ("LA", ["L", "X", "X", "A"]),
+        ("RGB", ["R", "G", "B", "X"]),
+        ("RGBX", ["R", "G", "B", "X"]),
+        ("RGBA", ["R", "G", "B", "A"]),
+        ("CMYK", ["C", "M", "Y", "K"]),
+        ("YCbCr", ["Y", "Cb", "Cr", "X"]),
+        ("HSV", ["H", "S", "V", "X"]),
+    ),
+)
+def test_image_metadata(mode: str, metadata: list[str]) -> None:
+    img = hopper(mode)
+
+    arr = Array(img)  # type: ignore[call-overload]
+
+    assert arr.type.value_field.metadata
+    assert arr.type.value_field.metadata[b"image"]
+
+    parsed_metadata = json.loads(arr.type.value_field.metadata[b"image"].decode("utf8"))
+
+    assert "bands" in parsed_metadata
+    assert parsed_metadata["bands"] == metadata

Tests/test_nanoarrow.py

@@ -0,0 +1,310 @@
+from __future__ import annotations
+
+import json
+from typing import Any, NamedTuple
+
+import pytest
+
+from PIL import Image
+
+from .helper import (
+    assert_deep_equal,
+    assert_image_equal,
+    hopper,
+    is_big_endian,
+)
+
+TYPE_CHECKING = False
+if TYPE_CHECKING:
+    import nanoarrow
+else:
+    nanoarrow = pytest.importorskip("nanoarrow", reason="Nanoarrow not installed")
+
+TEST_IMAGE_SIZE = (10, 10)
+
+
+def _test_img_equals_pyarray(
+    img: Image.Image, arr: Any, mask: list[int] | None, elts_per_pixel: int = 1
+) -> None:
+    assert img.height * img.width * elts_per_pixel == len(arr)
+    px = img.load()
+    assert px is not None
+    if elts_per_pixel > 1 and mask is None:
+        # have to do element-wise comparison when we're comparing
+        # flattened r,g,b,a to a pixel.
+        mask = list(range(elts_per_pixel))
+    for x in range(0, img.size[0], int(img.size[0] / 10)):
+        for y in range(0, img.size[1], int(img.size[1] / 10)):
+            if mask:
+                pixel = px[x, y]
+                assert isinstance(pixel, tuple)
+                for ix, elt in enumerate(mask):
+                    if elts_per_pixel == 1:
+                        assert pixel[ix] == arr[y * img.width + x].as_py()[elt]
+                    else:
+                        assert (
+                            pixel[ix]
+                            == arr[(y * img.width + x) * elts_per_pixel + elt].as_py()
+                        )
+            else:
+                assert_deep_equal(px[x, y], arr[y * img.width + x].as_py())
+
+
+def _test_img_equals_int32_pyarray(
+    img: Image.Image, arr: Any, mask: list[int] | None, elts_per_pixel: int = 1
+) -> None:
+    assert img.height * img.width * elts_per_pixel == len(arr)
+    px = img.load()
+    assert px is not None
+    if mask is None:
+        # have to do element-wise comparison when we're comparing
+        # flattened rgba in an uint32 to a pixel.
+        mask = list(range(elts_per_pixel))
+    for x in range(0, img.size[0], int(img.size[0] / 10)):
+        for y in range(0, img.size[1], int(img.size[1] / 10)):
+            pixel = px[x, y]
+            assert isinstance(pixel, tuple)
+            arr_pixel_int = arr[y * img.width + x].as_py()
+            arr_pixel_tuple = (
+                arr_pixel_int % 256,
+                (arr_pixel_int // 256) % 256,
+                (arr_pixel_int // 256**2) % 256,
+                (arr_pixel_int // 256**3),
+            )
+            if is_big_endian():
+                arr_pixel_tuple = arr_pixel_tuple[::-1]
+
+            for ix, elt in enumerate(mask):
+                assert pixel[ix] == arr_pixel_tuple[elt]
+
+
+fl_uint8_4_type = nanoarrow.fixed_size_list(
+    value_type=nanoarrow.uint8(nullable=False), list_size=4, nullable=False
+)
+
+
+@pytest.mark.parametrize(
+    "mode, dtype, mask",
+    (
+        ("L", nanoarrow.uint8(nullable=False), None),
+        ("I", nanoarrow.int32(nullable=False), None),
+        ("F", nanoarrow.float32(nullable=False), None),
+        ("LA", fl_uint8_4_type, [0, 3]),
+        ("RGB", fl_uint8_4_type, [0, 1, 2]),
+        ("RGBA", fl_uint8_4_type, None),
+        ("RGBX", fl_uint8_4_type, None),
+        ("CMYK", fl_uint8_4_type, None),
+        ("YCbCr", fl_uint8_4_type, [0, 1, 2]),
+        ("HSV", fl_uint8_4_type, [0, 1, 2]),
+    ),
+)
+def test_to_array(mode: str, dtype: nanoarrow, mask: list[int] | None) -> None:
+    img = hopper(mode)
+
+    # Resize to non-square
+    img = img.crop((3, 0, 124, 127))
+    assert img.size == (121, 127)
+
+    arr = nanoarrow.Array(img)  # type: ignore[call-overload]
+    _test_img_equals_pyarray(img, arr, mask)
+    assert arr.schema.type == dtype.type
+    assert arr.schema.nullable == dtype.nullable
+
+    reloaded = Image.fromarrow(arr, mode, img.size)
+
+    assert reloaded
+
+    assert_image_equal(img, reloaded)
+
+
+def test_lifetime() -> None:
+    # valgrind shouldn't error out here.
+    # arrays should be accessible after the image is deleted.
+
+    img = hopper("L")
+
+    arr_1 = nanoarrow.Array(img)  # type: ignore[call-overload]
+    arr_2 = nanoarrow.Array(img)  # type: ignore[call-overload]
+
+    del img
+
+    assert sum(arr_1.iter_py()) > 0
+    del arr_1
+
+    assert sum(arr_2.iter_py()) > 0
+    del arr_2
+
+
+def test_lifetime2() -> None:
+    # valgrind shouldn't error out here.
+    # img should remain after the arrays are collected.
+
+    img = hopper("L")
+
+    arr_1 = nanoarrow.Array(img)  # type: ignore[call-overload]
+    arr_2 = nanoarrow.Array(img)  # type: ignore[call-overload]
+
+    assert sum(arr_1.iter_py()) > 0
+    del arr_1
+
+    assert sum(arr_2.iter_py()) > 0
+    del arr_2
+
+    img2 = img.copy()
+    px = img2.load()
+    assert px  # make mypy happy
+    assert isinstance(px[0, 0], int)
+
+
+class DataShape(NamedTuple):
+    dtype: nanoarrow
+    # Strictly speaking, elt should be a pixel or pixel component, so
+    # list[uint8][4], float, int, uint32, uint8, etc.  But more
+    # correctly, it should be exactly the dtype from the line above.
+    elt: Any
+    elts_per_pixel: int
+
+
+UINT_ARR = DataShape(
+    dtype=fl_uint8_4_type,
+    elt=[1, 2, 3, 4],  # array of 4 uint8 per pixel
+    elts_per_pixel=1,  # only one array per pixel
+)
+
+UINT = DataShape(
+    dtype=nanoarrow.uint8(),
+    elt=3,  # one uint8,
+    elts_per_pixel=4,  # but repeated 4x per pixel
+)
+
+UINT32 = DataShape(
+    dtype=nanoarrow.uint32(),
+    elt=0xABCDEF45,  # one packed int, doesn't fit in a int32 > 0x80000000
+    elts_per_pixel=1,  # one per pixel
+)
+
+INT32 = DataShape(
+    dtype=nanoarrow.uint32(),
+    elt=0x12CDEF45,  # one packed int
+    elts_per_pixel=1,  # one per pixel
+)
+
+
+@pytest.mark.parametrize(
+    "mode, data_tp, mask",
+    (
+        ("L", DataShape(nanoarrow.uint8(), 3, 1), None),
+        ("I", DataShape(nanoarrow.int32(), 1 << 24, 1), None),
+        ("F", DataShape(nanoarrow.float32(), 3.14159, 1), None),
+        ("LA", UINT_ARR, [0, 3]),
+        ("LA", UINT, [0, 3]),
+        ("RGB", UINT_ARR, [0, 1, 2]),
+        ("RGBA", UINT_ARR, None),
+        ("CMYK", UINT_ARR, None),
+        ("YCbCr", UINT_ARR, [0, 1, 2]),
+        ("HSV", UINT_ARR, [0, 1, 2]),
+        ("RGB", UINT, [0, 1, 2]),
+        ("RGBA", UINT, None),
+        ("CMYK", UINT, None),
+        ("YCbCr", UINT, [0, 1, 2]),
+        ("HSV", UINT, [0, 1, 2]),
+    ),
+)
+def test_fromarray(mode: str, data_tp: DataShape, mask: list[int] | None) -> None:
+    (dtype, elt, elts_per_pixel) = data_tp
+
+    ct_pixels = TEST_IMAGE_SIZE[0] * TEST_IMAGE_SIZE[1]
+    if dtype == fl_uint8_4_type:
+        tmp_arr = nanoarrow.Array(
+            elt * (ct_pixels * elts_per_pixel), schema=nanoarrow.uint8()
+        )
+        c_array = nanoarrow.c_array_from_buffers(
+            dtype, ct_pixels, buffers=[], children=[tmp_arr]
+        )
+        arr = nanoarrow.Array(c_array)
+    else:
+        arr = nanoarrow.Array(
+            nanoarrow.c_array([elt] * (ct_pixels * elts_per_pixel), schema=dtype)
+        )
+    img = Image.fromarrow(arr, mode, TEST_IMAGE_SIZE)
+
+    _test_img_equals_pyarray(img, arr, mask, elts_per_pixel)
+
+
+@pytest.mark.parametrize(
+    "mode, data_tp, mask",
+    (
+        ("LA", UINT32, [0, 3]),
+        ("RGB", UINT32, [0, 1, 2]),
+        ("RGBA", UINT32, None),
+        ("CMYK", UINT32, None),
+        ("YCbCr", UINT32, [0, 1, 2]),
+        ("HSV", UINT32, [0, 1, 2]),
+        ("LA", INT32, [0, 3]),
+        ("RGB", INT32, [0, 1, 2]),
+        ("RGBA", INT32, None),
+        ("CMYK", INT32, None),
+        ("YCbCr", INT32, [0, 1, 2]),
+        ("HSV", INT32, [0, 1, 2]),
+    ),
+)
+def test_from_int32array(mode: str, data_tp: DataShape, mask: list[int] | None) -> None:
+    (dtype, elt, elts_per_pixel) = data_tp
+
+    ct_pixels = TEST_IMAGE_SIZE[0] * TEST_IMAGE_SIZE[1]
+    arr = nanoarrow.Array(
+        nanoarrow.c_array([elt] * (ct_pixels * elts_per_pixel), schema=dtype)
+    )
+    img = Image.fromarrow(arr, mode, TEST_IMAGE_SIZE)
+
+    _test_img_equals_int32_pyarray(img, arr, mask, elts_per_pixel)
+
+
+@pytest.mark.parametrize(
+    "mode, metadata",
+    (
+        ("LA", ["L", "X", "X", "A"]),
+        ("RGB", ["R", "G", "B", "X"]),
+        ("RGBX", ["R", "G", "B", "X"]),
+        ("RGBA", ["R", "G", "B", "A"]),
+        ("CMYK", ["C", "M", "Y", "K"]),
+        ("YCbCr", ["Y", "Cb", "Cr", "X"]),
+        ("HSV", ["H", "S", "V", "X"]),
+    ),
+)
+def test_image_nested_metadata(mode: str, metadata: list[str]) -> None:
+    img = hopper(mode)
+
+    arr = nanoarrow.Array(img)  # type: ignore[call-overload]
+
+    assert arr.schema.value_type.metadata
+    assert arr.schema.value_type.metadata[b"image"]
+
+    parsed_metadata = json.loads(
+        arr.schema.value_type.metadata[b"image"].decode("utf8")
+    )
+
+    assert "bands" in parsed_metadata
+    assert parsed_metadata["bands"] == metadata
+
+
+@pytest.mark.parametrize(
+    "mode, metadata",
+    (
+        ("L", ["L"]),
+        ("I", ["I"]),
+        ("F", ["F"]),
+    ),
+)
+def test_image_flat_metadata(mode: str, metadata: list[str]) -> None:
+    img = hopper(mode)
+
+    arr = nanoarrow.Array(img)  # type: ignore[call-overload]
+
+    assert arr.schema.metadata
+    assert arr.schema.metadata[b"image"]
+
+    parsed_metadata = json.loads(arr.schema.metadata[b"image"].decode("utf8"))
+
+    assert "bands" in parsed_metadata
+    assert parsed_metadata["bands"] == metadata

Tests/test_pyarrow.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import json
 from typing import Any, NamedTuple
 
 import pytest
@@ -244,3 +245,29 @@ def test_from_int32array(mode: str, data_tp: DataShape, mask: list[int] | None)
     img = Image.fromarrow(arr, mode, TEST_IMAGE_SIZE)
 
     _test_img_equals_int32_pyarray(img, arr, mask, elts_per_pixel)
+
+
+@pytest.mark.parametrize(
+    "mode, metadata",
+    (
+        ("LA", ["L", "X", "X", "A"]),
+        ("RGB", ["R", "G", "B", "X"]),
+        ("RGBX", ["R", "G", "B", "X"]),
+        ("RGBA", ["R", "G", "B", "A"]),
+        ("CMYK", ["C", "M", "Y", "K"]),
+        ("YCbCr", ["Y", "Cb", "Cr", "X"]),
+        ("HSV", ["H", "S", "V", "X"]),
+    ),
+)
+def test_image_metadata(mode: str, metadata: list[str]) -> None:
+    img = hopper(mode)
+
+    arr = pyarrow.array(img)  # type: ignore[call-overload]
+
+    assert arr.type.field(0).metadata
+    assert arr.type.field(0).metadata[b"image"]
+
+    parsed_metadata = json.loads(arr.type.field(0).metadata[b"image"].decode("utf8"))
+
+    assert "bands" in parsed_metadata
+    assert parsed_metadata["bands"] == metadata

pyproject.toml

@@ -57,6 +57,9 @@ optional-dependencies.mic = [
   "olefile",
 ]
 optional-dependencies.test-arrow = [
+  "arro3-compute",
+  "arro3-core",
+  "nanoarrow",
   "pyarrow",
 ]
 

src/libImaging/Arrow.c

@@ -55,6 +55,98 @@ ReleaseExportedSchema(struct ArrowSchema *array) {
     // Mark array released
     array->release = NULL;
 }
+char *
+image_band_json(Imaging im) {
+    char *format = "{\"bands\": [\"%s\", \"%s\", \"%s\", \"%s\"]}";
+    char *json;
+    // Bands can be 4 bands * 2 characters each
+    int len = strlen(format) + 8 + 1;
+    int err;
+
+    json = calloc(1, len);
+
+    if (!json) {
+        return NULL;
+    }
+
+    err = PyOS_snprintf(
+        json,
+        len,
+        format,
+        im->band_names[0],
+        im->band_names[1],
+        im->band_names[2],
+        im->band_names[3]
+    );
+    if (err < 0) {
+        return NULL;
+    }
+    return json;
+}
+
+char *
+single_band_json(Imaging im) {
+    char *format = "{\"bands\": [\"%s\"]}";
+    char *json;
+    // Bands can be 1 band * (maybe but probably not) 2 characters each
+    int len = strlen(format) + 2 + 1;
+    int err;
+
+    json = calloc(1, len);
+
+    if (!json) {
+        return NULL;
+    }
+
+    err = PyOS_snprintf(json, len, format, im->band_names[0]);
+    if (err < 0) {
+        return NULL;
+    }
+    return json;
+}
+
+char *
+assemble_metadata(const char *band_json) {
+    /* format is
+       int32: number of key/value pairs (noted N below)
+       int32: byte length of key 0
+       key 0 (not null-terminated)
+       int32: byte length of value 0
+       value 0 (not null-terminated)
+       ...
+       int32: byte length of key N - 1
+       key N - 1 (not null-terminated)
+       int32: byte length of value N - 1
+       value N - 1 (not null-terminated)
+    */
+    const char *key = "image";
+    INT32 key_len = strlen(key);
+    INT32 band_json_len = strlen(band_json);
+
+    char *buf;
+    INT32 *dest_int;
+    char *dest;
+
+    buf = calloc(1, key_len + band_json_len + 4 + 1 * 8);
+    if (!buf) {
+        return NULL;
+    }
+
+    dest_int = (void *)buf;
+
+    dest_int[0] = 1;
+    dest_int[1] = key_len;
+    dest_int += 2;
+    dest = (void *)dest_int;
+    memcpy(dest, key, key_len);
+    dest += key_len;
+    dest_int = (void *)dest;
+    dest_int[0] = band_json_len;
+    dest_int += 1;
+    memcpy(dest_int, band_json, band_json_len);
+
+    return buf;
+}
 
 int
 export_named_type(struct ArrowSchema *schema, char *format, char *name) {
@@ -95,6 +187,8 @@ export_named_type(struct ArrowSchema *schema, char *format, char *name) {
 int
 export_imaging_schema(Imaging im, struct ArrowSchema *schema) {
     int retval = 0;
+    char *metadata;
+    char *band_json;
 
     if (strcmp(im->arrow_band_format, "") == 0) {
         return IMAGING_ARROW_INCOMPATIBLE_MODE;
@@ -106,7 +200,17 @@ export_imaging_schema(Imaging im, struct ArrowSchema *schema) {
     }
 
     if (im->bands == 1) {
-        return export_named_type(schema, im->arrow_band_format, im->band_names[0]);
+        retval = export_named_type(schema, im->arrow_band_format, im->band_names[0]);
+        if (retval != 0) {
+            return retval;
+        }
+        // band related metadata
+        band_json = single_band_json(im);
+        if (band_json) {
+            schema->metadata = assemble_metadata(band_json);
+            free(band_json);
+        }
+        return retval;
     }
 
     retval = export_named_type(schema, "+w:4", "");
@@ -117,13 +221,24 @@ export_imaging_schema(Imaging im, struct ArrowSchema *schema) {
     schema->n_children = 1;
     schema->children = calloc(1, sizeof(struct ArrowSchema *));
     schema->children[0] = (struct ArrowSchema *)calloc(1, sizeof(struct ArrowSchema));
-    retval = export_named_type(schema->children[0], im->arrow_band_format, "pixel");
+    retval = export_named_type(schema->children[0], im->arrow_band_format, im->mode);
     if (retval != 0) {
         free(schema->children[0]);
         free(schema->children);
         schema->release(schema);
         return retval;
     }
+
+    // band related metadata
+    band_json = image_band_json(im);
+    if (band_json) {
+        // adding the metadata to the child array.
+        // Accessible in pyarrow via pa.array(img).type.field(0).metadata
+        // adding it to the top level is not accessible.
+        schema->children[0]->metadata = assemble_metadata(band_json);
+        free(band_json);
+    }
+
     return 0;
 }