"""色特徴量の抽出（`data/color_features.py`）のテスト. 仕様（実装サマリーの設計を仕様として扱う）: - `extract_color_features` は 107 次元（L*・a*・b*・C*・h ごとに平均・標準偏差・10/25/50/75/90 分位点 = 35，a*-b* 8×8 ヒストグラム = 64，L* 8 ビン = 8）．並びは `color_feature_names` - h は atan2(b*, a*) の度（(-180, 180]） - ヒストグラムは割合（合計 1）．範囲（a* [0, 50]・b* [-10, 30]・L* [20, 100]）外は端のビン - 舌領域（RGB がすべて 0 でない画素）だけを使う（背景の量・位置で特徴が変わらない） - 舌領域が 0 画素なら ValueError（ファイル名は出さない） - `downscale_nearest` は長辺が max_side を超えるときだけ最近傍法で縮小する（色が変わらない） """ from __future__ import annotations import numpy as np import pytest from tias_judge_model.data.color import rgb8_to_lab from tias_judge_model.data.color_features import ( NUM_COLOR_FEATURES, color_feature_names, downscale_nearest, extract_color_features, ) NAMES = color_feature_names() AB_BINS = 8 L_BINS = 8 A_RANGE = (0.0, 50.0) B_RANGE = (-10.0, 30.0) L_RANGE = (20.0, 100.0) def _idx(name: str) -> int: return NAMES.index(name) def _ab_hist(features: np.ndarray) -> np.ndarray: return np.array( [features[_idx(f"ab_hist_{i}_{j}")] for i in range(AB_BINS) for j in range(AB_BINS)] ).reshape(AB_BINS, AB_BINS) def _l_hist(features: np.ndarray) -> np.ndarray: return np.array([features[_idx(f"L_hist_{k}")] for k in range(L_BINS)]) def _uniform(rgb: tuple[int, int, int], shape: tuple[int, int] = (6, 5)) -> np.ndarray: img = np.zeros((*shape, 3), dtype=np.uint8) img[:] = rgb return img def _lab(rgb: tuple[int, int, int]) -> np.ndarray: return rgb8_to_lab(np.array([rgb], dtype=np.uint8))[0] def _random_tongue(rng: np.random.Generator, shape: tuple[int, int]) -> np.ndarray: # 1〜255 の値だけにして，すべての画素を舌領域にする return rng.integers(1, 256, size=(*shape, 3), dtype=np.uint8) # --- 次元・名前 --- def test_dimension_is_107() -> None: assert NUM_COLOR_FEATURES == 107 assert len(NAMES) == 107 rng = np.random.default_rng(0) f = extract_color_features(_random_tongue(rng, (8, 9))) assert f.shape == (107,) assert f.dtype == np.float64 assert np.all(np.isfinite(f)) def test_feature_names_are_unique_and_follow_layout() -> None: assert len(set(NAMES)) == len(NAMES) stats = ["mean", "std", "p10", "p25", "p50", "p75", "p90"] expected_channel = [f"{ch}_{s}" for ch in ("L", "a", "b", "C", "h") for s in stats] assert list(NAMES[:35]) == expected_channel assert list(NAMES[35:99]) == [f"ab_hist_{i}_{j}" for i in range(8) for j in range(8)] assert list(NAMES[99:]) == [f"L_hist_{k}" for k in range(8)] # --- 統計量 --- def test_uniform_color_statistics_match_lab() -> None: rgb = (225, 140, 140) f = extract_color_features(_uniform(rgb)) lightness, a, b = _lab(rgb) chroma = np.hypot(a, b) hue = np.degrees(np.arctan2(b, a)) for ch, value in zip("LabCh", (lightness, a, b, chroma, hue), strict=True): assert f[_idx(f"{ch}_mean")] == pytest.approx(value, abs=1e-6) assert f[_idx(f"{ch}_std")] == pytest.approx(0.0, abs=1e-6) for q in (10, 25, 50, 75, 90): assert f[_idx(f"{ch}_p{q}")] == pytest.approx(value, abs=1e-6) def test_mean_and_median_over_mixed_pixels() -> None: # 9 画素が色 B，1 画素が色 A: 平均は重み付き，中央値・p90 は B img = np.zeros((1, 10, 3), dtype=np.uint8) color_a, color_b = (120, 20, 30), (225, 140, 140) img[0, 0] = color_a img[0, 1:] = color_b f = extract_color_features(img) la, lb = _lab(color_a)[0], _lab(color_b)[0] assert f[_idx("L_mean")] == pytest.approx((la + 9 * lb) / 10, abs=1e-6) assert f[_idx("L_p50")] == pytest.approx(lb, abs=1e-6) assert f[_idx("L_p90")] == pytest.approx(lb, abs=1e-6) assert f[_idx("L_std")] > 0 # 分位点は単調非減少 ps = [f[_idx(f"L_p{q}")] for q in (10, 25, 50, 75, 90)] assert ps == sorted(ps) @pytest.mark.parametrize( "rgb", [(225, 140, 140), (110, 50, 140), (0, 255, 0), (0, 0, 255), (40, 200, 220)], ids=str, ) def test_hue_is_atan2_degrees_in_range(rgb: tuple[int, int, int]) -> None: f = extract_color_features(_uniform(rgb)) _, a, b = _lab(rgb) h = f[_idx("h_mean")] assert -180.0 < h <= 180.0 assert h == pytest.approx(np.degrees(np.arctan2(b, a)), abs=1e-6) def test_hue_sign_follows_b() -> None: # 紫（a* > 0, b* < 0）は負の色相，淡紅（a* > 0, b* > 0）は 0〜90° purple = extract_color_features(_uniform((110, 50, 140)))[_idx("h_mean")] pink = extract_color_features(_uniform((225, 140, 140)))[_idx("h_mean")] assert purple < 0 assert 0 < pink < 90 # --- ヒストグラム --- def test_histograms_sum_to_one_and_are_proportions() -> None: rng = np.random.default_rng(1) img = _random_tongue(rng, (13, 17)) img[:3] = 0 # 背景 f = extract_color_features(img) ab = _ab_hist(f) lh = _l_hist(f) assert ab.sum() == pytest.approx(1.0) assert lh.sum() == pytest.approx(1.0) assert np.all(ab >= 0) and np.all(lh >= 0) # 割合は舌領域の画素数の倍数 n = 10 * 17 assert np.allclose(ab * n, np.round(ab * n)) assert np.allclose(lh * n, np.round(lh * n)) def test_histogram_proportions_three_to_one() -> None: img = np.zeros((2, 3, 3), dtype=np.uint8) # 2 画素は背景 img[0, :3] = (225, 140, 140) img[1, 0] = (255, 255, 255) f = extract_color_features(img) lh = _l_hist(f) assert lh[7] == pytest.approx(0.25) # 白（L* = 100）は最上位ビン assert sorted(lh[lh > 0].tolist()) == pytest.approx([0.25, 0.75]) def test_in_range_color_goes_to_expected_ab_bin() -> None: rgb = (225, 140, 140) lightness, a, b = _lab(rgb) assert A_RANGE[0] < a < A_RANGE[1] and B_RANGE[0] < b < B_RANGE[1] i = int((a - A_RANGE[0]) // ((A_RANGE[1] - A_RANGE[0]) / AB_BINS)) j = int((b - B_RANGE[0]) // ((B_RANGE[1] - B_RANGE[0]) / AB_BINS)) k = int((lightness - L_RANGE[0]) // ((L_RANGE[1] - L_RANGE[0]) / L_BINS)) f = extract_color_features(_uniform(rgb)) ab = _ab_hist(f) assert ab[i, j] == pytest.approx(1.0) assert _l_hist(f)[k] == pytest.approx(1.0) @pytest.mark.parametrize( ("rgb", "ab_bin"), [ ((255, 0, 0), (7, 7)), # a* ≈ 80・b* ≈ 67（両方上限超え） ((0, 255, 0), (0, 7)), # a* ≈ -86（下限未満）・b* ≈ 83（上限超え） ((0, 0, 255), (7, 0)), # a* ≈ 79（上限超え）・b* ≈ -108（下限未満） ], ids=["red", "green", "blue"], ) def test_out_of_range_ab_goes_to_edge_bin( rgb: tuple[int, int, int], ab_bin: tuple[int, int] ) -> None: f = extract_color_features(_uniform(rgb)) ab = _ab_hist(f) assert ab[ab_bin] == pytest.approx(1.0) assert ab.sum() == pytest.approx(1.0) def test_out_of_range_lightness_goes_to_edge_bin() -> None: img = np.zeros((1, 4, 3), dtype=np.uint8) img[0, 0] = (5, 5, 5) # L* ≈ 1.4（下限 20 未満）→ 最下位ビン img[0, 1] = (3, 3, 3) img[0, 2] = (255, 255, 255) # L* = 100（上限ちょうど）→ 最上位ビン img[0, 3] = (255, 255, 255) lh = _l_hist(extract_color_features(img)) assert lh[0] == pytest.approx(0.5) assert lh[7] == pytest.approx(0.5) assert lh.sum() == pytest.approx(1.0) # --- 背景の扱い --- def test_background_amount_and_position_do_not_change_features() -> None: rng = np.random.default_rng(2) tongue = _random_tongue(rng, (10, 12)) small = tongue.copy() large = np.zeros((40, 55, 3), dtype=np.uint8) large[17:27, 30:42] = tongue np.testing.assert_allclose( extract_color_features(small), extract_color_features(large), rtol=0, atol=1e-9 ) def test_black_holes_inside_tongue_are_ignored() -> None: rng = np.random.default_rng(3) tongue = _random_tongue(rng, (1, 20)) with_holes = np.zeros((3, 20, 3), dtype=np.uint8) with_holes[1] = tongue # 舌の画素の順序を変えても（並べ替え不変），黒を挟んでも同じ shuffled = np.zeros((20, 2, 3), dtype=np.uint8) shuffled[:, 0] = tongue[0][::-1] np.testing.assert_allclose( extract_color_features(tongue), extract_color_features(with_holes), atol=1e-9 ) np.testing.assert_allclose( extract_color_features(tongue), extract_color_features(shuffled), atol=1e-9 ) def test_nearly_black_pixel_counts_as_tongue() -> None: # RGB がすべて 0 でなければ舌領域（1 チャンネルだけ 1 でも数える） img = np.zeros((3, 3, 3), dtype=np.uint8) img[1, 1] = (0, 0, 1) f = extract_color_features(img) assert _l_hist(f).sum() == pytest.approx(1.0) assert f[_idx("L_std")] == pytest.approx(0.0, abs=1e-9) def test_no_tongue_pixels_raises_without_file_name() -> None: with pytest.raises(ValueError) as excinfo: extract_color_features(np.zeros((5, 5, 3), dtype=np.uint8)) msg = str(excinfo.value) assert ".png" not in msg and ".jpg" not in msg @pytest.mark.parametrize("shape", [(5, 5), (5, 5, 4), (5, 5, 1)], ids=str) def test_invalid_shape_raises(shape: tuple[int, ...]) -> None: with pytest.raises(ValueError): extract_color_features(np.full(shape, 100, dtype=np.uint8)) # --- 最近傍縮小 --- def test_downscale_keeps_small_images_unchanged() -> None: rng = np.random.default_rng(4) img = _random_tongue(rng, (30, 50)) np.testing.assert_array_equal(downscale_nearest(img, 50), img) np.testing.assert_array_equal(downscale_nearest(img, 100), img) @pytest.mark.parametrize( ("shape", "max_side", "expected"), [((40, 100), 50, (20, 50)), ((100, 30), 50, (50, 15)), ((64, 64), 16, (16, 16))], ) def test_downscale_limits_long_side_keeping_aspect( shape: tuple[int, int], max_side: int, expected: tuple[int, int] ) -> None: rng = np.random.default_rng(5) out = downscale_nearest(_random_tongue(rng, shape), max_side) assert out.shape == (*expected, 3) assert out.dtype == np.uint8 assert max(out.shape[:2]) <= max_side def test_downscale_does_not_create_new_colors() -> None: rng = np.random.default_rng(6) img = np.zeros((90, 70, 3), dtype=np.uint8) img[10:80, 10:60] = _random_tongue(rng, (70, 50)) out = downscale_nearest(img, 23) src_colors = {tuple(p) for p in img.reshape(-1, 3)} out_colors = {tuple(p) for p in out.reshape(-1, 3)} assert out_colors <= src_colors def test_downscale_of_block_image_keeps_features() -> None: # 2×2 の一様なブロックからなる画像を 1/2 に縮小しても，色の割合は変わらない rng = np.random.default_rng(7) small = _random_tongue(rng, (16, 24)) small[:4, :4] = 0 # 背景も含める big = np.repeat(np.repeat(small, 2, axis=0), 2, axis=1) # (32, 48, 3) out = downscale_nearest(big, 24) assert out.shape == (16, 24, 3) # 縮小後は各ブロックの色がちょうど 1 画素ずつ残る（色が混ざらず，割合も保たれる） np.testing.assert_array_equal(out, small) # 分位点は線形補間のため重複度で僅かに変わりうるので，平均・ヒストグラムで比べる keep = [i for i, n in enumerate(NAMES) if not n.split("_")[-1].startswith("p")] f_out, f_big = extract_color_features(out), extract_color_features(big) np.testing.assert_allclose(f_out[keep], f_big[keep], atol=1e-9) @pytest.mark.parametrize("max_side", [0, -1]) def test_downscale_rejects_non_positive_max_side(max_side: int) -> None: with pytest.raises(ValueError): downscale_nearest(np.ones((4, 4, 3), dtype=np.uint8), max_side)