Peach: Local-first video metadata scraper and media manager.
fix(covers): 重算过期人脸缓存并保留整幅主脸
longmeidaoef89e39ad002
diff --git a/docs/REUSE.md b/docs/REUSE.mdindex e02ecba04..3705fd488 100644--- a/docs/REUSE.md+++ b/docs/REUSE.md@@ -217,6 +217,8 @@ JAV 默认封面(官方封面/预览图)与视频默认大小(大图/无快照的有码作品由封面入口复用现有 `DmmSource` 的 GraphQL 查询与精确身份核验;成功快照直接展开 DMM 图片候选,r18.dev 与 DMM 使用同一份目录证据。FC2 来源快照通过 `storage_original` 提取发行方存储原件,排除社区图床。目录查询沿用来源冷却,HTTP、网络与解析失败分别保留未取得状态。+人脸缓存使用 `cover-face-v2` 与源图尺寸校验。`detect_cover_faces.py` 自动重算过期记录,逐图核查加 `--refresh-faces` 可在重算整图主脸后再决定版式;无可信书脊且左侧主脸明显占优的大封按整幅画面取景。安装、批次重算与恢复默认取景读取同一版式标记。+取景由 `poster-crop-v12` 管理。Blu-ray 来源标题、明确的版次标签与图片比例共同决定 0.85 正封比例,标题随正封框保留,重算人脸与正封框均读取同一版次;有效宽正封框在没有版次标题的卡片上也完整显示。VR 标记、IENEE、IENFE、SDHS 数字发行拼图,以及可信主脸跨过 DVD 正封先验边界的大封保留整幅画面,由 `cover_frame.layout=still` 传给卡片。大封在竖版大图按主脸居中填满,横版小图完整显示,均不使用模糊补边;JPEG 软边的书脊识别结合突出度、左忙右静的方向与正封侧基线。手工框保持用户取景。重试队列按最早尝试时间轮转,收起的合集沿用设置跳过;逐番号结果保存在 `feed-cover-attempts/`。复用现有 HTTP、Pillow、OpenCV 与文件锁,无新增依赖。馆藏与新作取图共用 `cover_artwork.metadata_snapshots` 和 `best_cover`,片商官网、发行方、R18 与 DMM 的官方候选均参与尺寸比较。馆藏采集将长边 1600px 作为高清重探界限,社区转存图可由官方封面替换;同类官方图按像素面积升级,FC2 保留商品图的来源保护。`scripts/audit_cover_artwork.py` 逐张核对尺寸、来源、版次与取景框,输出 CSV;`scripts/upgrade_audited_covers.py` 按已有官方快照逐番号重探,`--live` 经同一官方链在线补查,保存原件备份与结果检查点。两者只处理可重建图片和边车,不写 ledger。diff --git a/docs/adr/0107-cover-quality-and-javdb-evidence.md b/docs/adr/0107-cover-quality-and-javdb-evidence.mdindex bfc9c95f9..edb4f5d15 100644--- a/docs/adr/0107-cover-quality-and-javdb-evidence.md+++ b/docs/adr/0107-cover-quality-and-javdb-evidence.md@@ -17,6 +17,7 @@- JavDB 图片只作来源复核与图像一致性证据,不作为正式封面、下载兜底或页面显示兜底。安装、封面索引与详情投影均执行此边界。官方图片未取得时显示未取得状态。- 有 JavDB 来源证据的历史封面及其边车移入 `review/` 下的可恢复归档,保留逐番号清单。来源未登记不等于 JavDB,不凭缺少来源批量移走文件。整个流程不写 ledger。- `poster-crop-v12` 结合主脸是否跨越正封先验边界识别整幅大封,结合梯度突出度与正封侧基线收紧书脊切点;手工取景保持原样。+- `cover-face-v2` 在整图上挑主脸,再决定是否按正封过滤。没有可辨书脊且左侧主脸面积至少为右侧两倍时按整幅大封取景;可信书脊保持传统封套布局。人脸缓存校验算法版本与图片尺寸,逐图核查支持 `--refresh-faces`。- 整幅大封在竖版大图卡片按主脸居中填满,在横版小图卡片完整放入容器;两种视图均不使用模糊补边。图片比例不同,横版完整显示时允许实色留边。## 验证与复核产物diff --git a/scripts/audit_cover_artwork.py b/scripts/audit_cover_artwork.pyindex 280333b91..2bacc610d 100644--- a/scripts/audit_cover_artwork.py+++ b/scripts/audit_cover_artwork.py@@ -13,7 +13,7 @@ sys.path.insert(0, str(Path(__file__).resolve().parents[1] / 'src'))from PIL import Imagefrom peach import jav_poster_crop as crop-from peach.cover_artwork import metadata_snapshots+from peach.cover_artwork import FACE_ALGORITHM_VERSION, metadata_snapshots, face_record as detect_facefrom peach.community_catalog import origin_offrom peach.fsutil import atomic_write_textfrom peach.library_processing import _official_evidence@@ -31,13 +31,20 @@ def snapshots(root, code):return metadata_snapshots(root, code)-def inspect(path, metadata):+def current_face(path, size, title, detector):+ """按算法版本与源图尺寸核对缓存,需要时从整图重新检测。"""+ record = read(path.with_suffix('.face.json'))+ if detector is not None and (record.get('version') != FACE_ALGORITHM_VERSION or record.get('px') != list(size)):+ from peach.face_detect import decode+ return detect_face(decode(path.read_bytes()), detector, code=path.stem, title=title) or {}+ return record+++def inspect(path, metadata, *, detector=None):from peach.library_processing import artwork_titlewith Image.open(path) as image:width, height = image.sizeimage.verify()- face_record = read(path.with_suffix('.face.json'))- face = face_record.get('face') if face_record.get('px') == [width, height] else Nonesaved = crop.read_sidecar(path) or {}evidence = read(path.with_suffix('.scraping.json'))known = snapshots(metadata, path.stem)@@ -45,10 +52,12 @@ def inspect(path, metadata):title = artwork_title([*known, ('installed', {'title': previous_title, 'genres': [previous_title]})])if crop.is_vr_title(previous_title):title = previous_title+ face_record = current_face(path, (width, height), title, detector)+ face = face_record.get('face') if face_record.get('px') == [width, height] else Nonefull = crop.face_crosses_spine(width, height, face)layout = 'still' if full or crop.is_scene_cover(path.stem, title) or face_record.get('layout') == 'still' else ''proposed = saved if crop.is_manual(saved) and crop.is_current(saved, width, height) else crop.crop_record(- path.stem, width, height, crop.file_gradient(path), title=title, face=face)+ path.stem, width, height, crop.file_gradient(path), title=title, face=face, layout=layout)if layout == 'still' and not crop.is_manual(proposed):proposed['box'] = {'x0': 0, 'y0': 0, 'x1': width, 'y1': height, 'method': crop.NONE}origin = origin_of(str(evidence.get('source_url') or ''))@@ -61,6 +70,8 @@ def inspect(path, metadata):reasons.append('low_resolution')if not crop.is_current(saved, width, height):reasons.append('stale_frame')+ if face_record.get('version') != FACE_ALGORITHM_VERSION:+ reasons.append('stale_face')if saved.get('box') != proposed.get('box'):reasons.append('crop_changed')if proposed['box']['method'] == crop.RATIO:@@ -80,13 +91,18 @@ def main():parser.add_argument('--data-root', type=Path, required=True)parser.add_argument('--report', type=Path, required=True)parser.add_argument('--apply', action='store_true', help='备份并更新可重建的取景边车')+ parser.add_argument('--refresh-faces', action='store_true', help='按当前人脸算法重算过期缓存')args = parser.parse_args()args.report.mkdir(parents=True, exist_ok=True)rows, errors = [], []covers = sorted((args.data_root / 'generated/covers').glob('*.jpg'))+ detector = None+ if args.refresh_faces:+ from peach.face_detect import FaceDetector+ detector = FaceDetector(allow_download=False)for index, path in enumerate(covers, 1):try:- row, poster, face, layout = inspect(path, args.data_root / 'sources/library-metadata')+ row, poster, face, layout = inspect(path, args.data_root / 'sources/library-metadata', detector=detector)rows.append(row)if args.apply:changes = {path.with_suffix('.poster.json'): poster}diff --git a/scripts/detect_cover_faces.py b/scripts/detect_cover_faces.pyindex bb136706e..d4dc8dc1c 100644--- a/scripts/detect_cover_faces.py+++ b/scripts/detect_cover_faces.py@@ -10,13 +10,11 @@ r"""给已落盘的封面算人脸位置,供卡片取景用。- 16:9 官方剧照在大图容器里被横向裁,整幅都是画面,没有「正封那一块」可推。横向锚点不跟着人走的话,偏在一侧的人物会整个落到可见窗口外面。-判据来自 46 张真实封面的实测:检出集中在长封套右侧正封的人脸位置(x 68~90%、-y 12~33%);封套最左的剧照拼贴里也有真脸(`278GYAN-017`、`KAVR-428`),所以-**长封套上落在左半边的检出一律丢弃**——那是拼贴区,不是正封。剧照没有拼贴区,-不套这条。检出不是取景的前提,取不到就退回固定值。+取景复用 `cover_artwork.face_record`:先检查整幅画面的主脸,再判断是否存在可信书脊。+传统封套使用右侧正封人脸,整幅大封使用主脸。记录包含算法版本和源图尺寸,过期缓存自动重算。结果写成每张封面一个 sidecar(`<番号>.face.json`),不写单一索引文件:抓取器-可能正在并发写同目录,单文件会互相覆盖。脚本可反复运行,已算过的默认跳过。+可能正在并发写同目录,单文件会互相覆盖。脚本可反复运行,当前版本且尺寸匹配的记录默认跳过。"""from __future__ import annotations@@ -31,7 +29,7 @@ if str(SRC_DIR) not in sys.path:sys.path.insert(0, str(SRC_DIR))from peach.config import COVER_DIR-from peach.cover_artwork import face_record+from peach.cover_artwork import FACE_ALGORITHM_VERSION, face_recordfrom peach.face_detect import FaceDetector, FaceModelUnavailablefrom peach.jav_poster_crop import read_sidecar@@ -50,6 +48,16 @@ def build_parser() -> argparse.ArgumentParser:return parser+def current_face(path):+ from peach.images import measure_image_file+ try:+ record = json.loads(path.with_suffix('.face.json').read_text(encoding='utf-8'))+ except (OSError, ValueError):+ return False+ return (isinstance(record, dict) and record.get('version') == FACE_ALGORITHM_VERSION+ and record.get('px') == list(measure_image_file(path) or ()))++def run(args: argparse.Namespace) -> int:try:detector = FaceDetector()@@ -58,7 +66,7 @@ def run(args: argparse.Namespace) -> int:covers = sorted(args.covers.glob("*.jpg"))todo = [p for p in covers- if args.redo or not p.with_suffix(".face.json").is_file()]+ if args.redo or not current_face(p)]if args.limit:todo = todo[:args.limit]print(f"封面 {len(covers)} 张,待处理 {len(todo)} 张")diff --git a/scripts/poster_crop_boxes.py b/scripts/poster_crop_boxes.pyindex b2764d94c..aa4c3396d 100644--- a/scripts/poster_crop_boxes.py+++ b/scripts/poster_crop_boxes.py@@ -101,7 +101,8 @@ def run(args: argparse.Namespace) -> int:record = jav_poster_crop.crop_record(path.stem, width, height, jav_poster_crop.file_gradient(path),title=(jav_poster_crop.read_sidecar(path) or {}).get('title', ''),- face=face.get('face') if face.get('px') == [width, height] else None)+ face=face.get('face') if face.get('px') == [width, height] else None,+ layout=face.get('layout', '') if face.get('px') == [width, height] else '')method = record["box"]["method"]counts[method] += 1bucket = samples.setdefault(method, [])diff --git a/src/peach/cover_artwork.py b/src/peach/cover_artwork.pyindex 6a8fbff19..798a1a75d 100644--- a/src/peach/cover_artwork.py+++ b/src/peach/cover_artwork.py@@ -16,6 +16,7 @@ from .jav_cover_fetch import Unavailable, is_javdb_imageFRONT_START = 0.468MIN_FACE_Y, MAX_FACE_Y = 0.05, 0.60+FACE_ALGORITHM_VERSION = 'cover-face-v2'_INSTALL_LOCK = threading.RLock()@@ -46,6 +47,20 @@ def metadata_snapshots(root: Path, code: str) -> list[tuple[str, dict]]:return found+def _dominant_left_face(image, largest, faces, *, code, title):+ """没有可辨书脊、左侧主脸明显大于右侧拼图时,保留整幅宣传画面。"""+ height, width = image.shape[:2]+ if not (jav_poster_crop.SLEEVE_RATIO_MIN <= width / height < jav_poster_crop.SLEEVE_RATIO_MAX):+ return False+ if largest is None or largest.cx >= FRONT_START or largest.score < .8 or largest.width < .075 or largest.height < .12:+ return False+ front = main_face([face for face in faces if face.cx >= FRONT_START])+ if front and largest.width * largest.height < 2 * front.width * front.height:+ return False+ box = jav_poster_crop.front_panel_box(width, height, lambda: jav_poster_crop.column_gradient(image), code, title)+ return box['method'] == jav_poster_crop.RATIO++def face_record(image, detector, *, code: str, title: str = '') -> dict | None:"""已解码的封面 → 页面取景记录;无法解码时不生成记录。@@ -58,10 +73,12 @@ def face_record(image, detector, *, code: str, title: str = '') -> dict | None:height, width = image.shape[:2]ratio = width / height if height else 0faces = detector.detect(image)- largest = main_face([face for face in faces if MIN_FACE_Y <= face.cy <= MAX_FACE_Y])+ portraits = [face for face in faces if MIN_FACE_Y <= face.cy <= MAX_FACE_Y]+ largest = main_face(portraits)full_cover = jav_poster_crop.face_crosses_spine(width, height, {'cx': largest.cx, 'w': largest.width, 'h': largest.height,- 'score': largest.score} if largest else None)+ 'score': largest.score} if largest else None) or _dominant_left_face(+ image, largest, portraits, code=code, title=title)if (not full_cover and jav_poster_crop.crops_to_portrait(code, title)and (jav_poster_crop.SLEEVE_RATIO_MIN <= ratio < jav_poster_crop.SLEEVE_RATIO_MAXor (jav_poster_crop.is_wide_sleeve_label(code, title)@@ -70,7 +87,7 @@ def face_record(image, detector, *, code: str, title: str = '') -> dict | None:low, high = (0.0, 1.0) if jav_poster_crop.is_scene_cover(code, title) else (MIN_FACE_Y, MAX_FACE_Y)faces = [face for face in faces if low <= face.cy <= high]best = main_face(faces)- return {"ratio": round(ratio, 3), "px": [int(width), int(height)],+ return {"version": FACE_ALGORITHM_VERSION, "ratio": round(ratio, 3), "px": [int(width), int(height)],**({'layout': 'still'} if full_cover or jav_poster_crop.is_scene_cover(code, title) else {}), "face": ({"cx": best.cx, "cy": best.cy, "w": best.width, "h": best.height,"score": best.score} if best else None)}@@ -91,7 +108,7 @@ def _sidecars(code: str, payload: bytes, size: tuple[int, int], detector=None, eposter = jav_poster_crop.crop_record(code, size[0], size[1],(lambda: jav_poster_crop.column_gradient(image)) if image is not None else None,- title=title, face=(face or {}).get('face'),+ title=title, face=(face or {}).get('face'), layout=(face or {}).get('layout', ''),)records = {jav_poster_crop.SIDECAR_SUFFIX: poster}if face is not None:diff --git a/src/peach/jav_poster_crop.py b/src/peach/jav_poster_crop.pyindex b21810dfa..941a33d1b 100644--- a/src/peach/jav_poster_crop.py+++ b/src/peach/jav_poster_crop.py@@ -566,14 +566,15 @@ def _panel(width: int, height: int, start: int, method: str) -> dict:def crop_record(code: str | None, width: int, height: int,- gradient_source: GradientSource = None, *, title: str = '', face: dict | None = None) -> dict:+ gradient_source: GradientSource = None, *, title: str = '', face: dict | None = None,+ layout: str = '') -> dict:"""可直接落盘的 sidecar 内容:算法版本、源图尺寸、框。番号形态先判:不该裁的图连梯度都不算,`gradient_source` 是惰性的就一次都不解码。"""width, height = int(width or 0), int(height or 0)box = (front_panel_box(width, height, gradient_source, code, title)- if crops_to_portrait(code, title) and not face_crosses_spine(width, height, face)+ if layout != 'still' and crops_to_portrait(code, title) and not face_crosses_spine(width, height, face)else _whole(width, height))return {"version": ALGORITHM_VERSION, "px": [width, height], "box": box,**({'title': title} if title else {})}diff --git a/tests/test_jav_cover_artwork.py b/tests/test_jav_cover_artwork.pyindex 913db7dc1..41b50958c 100644--- a/tests/test_jav_cover_artwork.py+++ b/tests/test_jav_cover_artwork.py@@ -60,6 +60,27 @@ class CoverArtworkTests(unittest.TestCase):self.assertEqual(face['layout'], 'still')self.assertIsNone(jav_poster_crop.projection(jav_poster_crop.read_sidecar(self.target)))+ def test_dominant_left_portrait_without_a_spine_keeps_the_whole_picture(self):+ detector = Detector()+ with patch.object(detector, 'detect', return_value=[Face(.22, .3, .3, .4, .95), Face(.8, .3, .06, .1, .9)]):+ cover_artwork.install_cover(self.target, 'MIDV-865', jpeg((80, 90, 100), (800, 600)),+ (800, 600), detector=detector)+ face = json.loads(self.target.with_suffix('.face.json').read_text(encoding='utf-8'))+ self.assertEqual((face['version'], face['layout'], face['face']['cx']),+ (cover_artwork.FACE_ALGORITHM_VERSION, 'still', .22))+ self.assertIsNone(jav_poster_crop.projection(jav_poster_crop.read_sidecar(self.target)))++ def test_a_clear_spine_keeps_the_front_portrait_despite_a_larger_back_portrait(self):+ detector = Detector()+ with patch.object(detector, 'detect', return_value=[Face(.22, .3, .3, .4, .95), Face(.8, .3, .06, .1, .9)]):+ with patch.object(jav_poster_crop, 'front_panel_box', return_value={+ 'x0': 420, 'y0': 0, 'x1': 800, 'y1': 600, 'method': 'fold'}):+ cover_artwork.install_cover(self.target, 'ABW-232', jpeg((80, 90, 100), (800, 600)),+ (800, 600), detector=detector)+ face = json.loads(self.target.with_suffix('.face.json').read_text(encoding='utf-8'))+ self.assertNotIn('layout', face)+ self.assertEqual(face['face']['cx'], .8)+def test_full_photo_keeps_faces_on_the_left_but_sleeves_use_the_front(self):detector = Detector()with patch.object(detector, "detect", return_value=[Face(.22, .3, .1, .1, .95)]):diff --git a/tests/test_jav_poster_crop.py b/tests/test_jav_poster_crop.pyindex 4ceee3a76..1dcf71dc9 100644--- a/tests/test_jav_poster_crop.py+++ b/tests/test_jav_poster_crop.py@@ -782,6 +782,31 @@ class BatchScriptTests(unittest.TestCase):self.assertEqual((proposed['box']['method'], layout), (NONE, 'still'))self.assertEqual(row['x0'], 0)+ def test_audit_refreshes_legacy_front_only_faces_before_classifying_the_cover(self):+ from peach.face_detect import Face+ from peach.cover_artwork import FACE_ALGORITHM_VERSION+ from unittest.mock import Mock+ target = self.root / 'SPONE-001.jpg'+ Image.new('RGB', (800, 600), 'gray').save(target)+ target.with_suffix('.face.json').write_text(json.dumps({+ 'px': [800, 600], 'face': {'cx': .8, 'cy': .3, 'w': .06, 'h': .1, 'score': .9},+ }), encoding='utf-8')+ detector = Mock()+ detector.detect.return_value = [Face(.555, .331, .257, .422, .918), Face(.8, .3, .06, .1, .9)]+ row, proposed, face, layout = batch_script('audit_cover_artwork').inspect(target, self.root / 'metadata', detector=detector)+ self.assertEqual((proposed['box']['method'], layout), (NONE, 'still'))+ self.assertEqual((face['version'], face['face']['cx']), (FACE_ALGORITHM_VERSION, .555))+ self.assertNotIn('stale_face', row['reasons'])++ def test_face_cache_requires_current_algorithm_and_image_dimensions(self):+ from peach.cover_artwork import FACE_ALGORITHM_VERSION+ script = batch_script('detect_cover_faces')+ target = self.cover.with_suffix('.face.json')+ for version, size, expected in (('old', [800, 540], False),+ (FACE_ALGORITHM_VERSION, [1, 1], False), (FACE_ALGORITHM_VERSION, [800, 540], True)):+ target.write_text(json.dumps({'version': version, 'px': size}), encoding='utf-8')+ self.assertEqual(script.current_face(self.cover), expected)+def test_apply_writes_the_sidecar_and_leaves_the_image_untouched(self):self.assertEqual(self.run_script("--apply"), 0)record = json.loads(