screenshot-vision-algorithm 0.3.8__tar.gz → 0.3.10__tar.gz

{screenshot_vision_algorithm-0.3.8 → screenshot_vision_algorithm-0.3.10}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: screenshot_vision_algorithm
-Version: 0.3.8
+Version: 0.3.10
 Summary: Multi-platform screenshot vision algorithms: WeChat (card bbox, speaker bands, avatar detection, OCR helpers), XHS & Zhihu (planned)
 License: MIT
 Requires-Python: >=3.10

{screenshot_vision_algorithm-0.3.8 → screenshot_vision_algorithm-0.3.10}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "screenshot_vision_algorithm"
-version = "0.3.8"
+version = "0.3.10"
 description = "Multi-platform screenshot vision algorithms: WeChat (card bbox, speaker bands, avatar detection, OCR helpers), XHS & Zhihu (planned)"
 readme = "README.md"
 license = { text = "MIT" }

{screenshot_vision_algorithm-0.3.8 → screenshot_vision_algorithm-0.3.10}/src/screenshot_vision_algorithm/android/wechat/algorithms/speaker_band.py

@@ -221,6 +221,7 @@ def build_prd_speaker_vertical_bands(
     scale_w: float,
     first_avatar_gap_reject_baseline: int = CHAT_BAND_FIRST_AVATAR_TOP_GAP_REJECT_BASELINE,
     first_band_top_extend_baseline: int = CHAT_FIRST_BAND_TOP_EXTEND_BASELINE,
+    avatar_merge_min_sep_baseline: int = 24,
 ) -> PrdChatVerticalLayout:
     """由时间序头像顶构造 PRD §5 顶隙 orphan + 左闭右开发言人筒。
 
@@ -248,7 +249,7 @@ def build_prd_speaker_vertical_bands(
 
     # 时间轴序去重：过近视为同一人头像重复检出
     merged_starts: list[int] = []
-    min_sep = int(round(24 * sw))
+    min_sep = int(round(int(avatar_merge_min_sep_baseline) * sw))
     for e in timeline:
         y = int(e.y_top)
         if not merged_starts or y - merged_starts[-1] >= min_sep:

{screenshot_vision_algorithm-0.3.8 → screenshot_vision_algorithm-0.3.10}/src/screenshot_vision_algorithm/android/wechat/ocr/nickname_binding.py

@@ -230,6 +230,32 @@ class NicknameOcrConfig:
     """Prefix / substring tags that exclude a block from nickname
     consideration even if all geometric criteria pass."""
 
+    # ── 以下字段可由 rules_config.yaml 覆盖（housekeeping 胶水层读 yaml → 灌参）──
+
+    skip_block_patterns: tuple[str, ...] = ()
+    """附加正则列表，逐条编译后传递给 ``_should_skip_block``。
+    内置正则（时间戳/纯数字/系统提示等）**始终并行生效**，
+    config 提供的是追加规则，不替代内置逻辑。"""
+
+    inside_y_tolerance_px: float = 4.0
+    """``_completely_inside`` 的 Y 方向容差（px @ 1080 基准 × scale_w）。
+    默认 4.0 与 ``card_binding.inside_y_tolerance_px`` 对齐。"""
+
+    cross_band_distance_threshold_px: float = 50.0
+    """Step C 跨 band 距离门控基线（px @ 1080 × scale_w）。
+    文字底部到卡片所属 speaker_band 顶部超过此值则不作为昵称。"""
+
+    # ── speaker_band 参数（同名传给 build_prd_speaker_vertical_bands）──
+
+    first_avatar_gap_reject_baseline: int = 30
+    """orphan 间隙阈值（px @ 1080）。"""
+
+    first_band_top_extend_baseline: int = 20
+    """首个 band 上移量（px @ 1080）。"""
+
+    avatar_merge_min_sep_px: int = 24
+    """头像合并最小间距（px @ 1080 × scale_w）。"""
+
 
 _ATTRIBUTION_PREFIX_DROPS: tuple[str, ...] = (
     "撤回了一条消息",
@@ -710,7 +736,12 @@ def _load_prd_layout_optional(
     bounds = compute_chat_content_vertical_bounds(arr, sw, screen_h=oh)
     ly, ry = detect_chat_side_avatar_ytops(arr, sw, bounds)
     timeline = merge_avatar_ytops_time_order(ly, ry)
-    return build_prd_speaker_vertical_bands(timeline, bounds, scale_w=sw)
+    return build_prd_speaker_vertical_bands(
+        timeline, bounds, scale_w=sw,
+        first_avatar_gap_reject_baseline=config.first_avatar_gap_reject_baseline,
+        first_band_top_extend_baseline=config.first_band_top_extend_baseline,
+        avatar_merge_min_sep_baseline=config.avatar_merge_min_sep_px,
+    )
 
 
 def _finalize_prd_attribution(
@@ -887,7 +918,7 @@ def _resume_thumb_bindings_and_orphans(
     # OCR bbox 与卡片 bbox 各有 ±2~3px 抖动：卡内标题行可能比卡顶高出 1~3px，
     # 若按严格包含判定会被误认为"卡外文字"而误绑为昵称（2026-06-11 真实案例）。
     # 真昵称行整体位于卡片上方（y2 < 卡顶），不受此容差影响。
-    _inside_tol = 4.0 * max(scale_w, 1.0)
+    _inside_tol = float(config.inside_y_tolerance_px) * max(scale_w, 1.0)
 
     def _completely_inside(block_bbox: tuple, card_bbox: tuple) -> bool:
         bx1, by1, bx2, by2 = block_bbox
@@ -917,8 +948,25 @@ def _resume_thumb_bindings_and_orphans(
             return True
         if _is_favorite_card_label(t):
             return True
+        # ── 附加正则（来自 rules_config.yaml skip_patterns，编译后缓存）──
+        if _compiled_skip_patterns:
+            for _pat in _compiled_skip_patterns:
+                try:
+                    if _pat.search(t):
+                        return True
+                except Exception:
+                    continue
         return False
 
+    # ── 编译 config 追加的正则（函数内缓存，避免每次 block 重新编译）──
+    _compiled_skip_patterns: list[re.Pattern] = []
+    if config.skip_block_patterns:
+        for p in config.skip_block_patterns:
+            try:
+                _compiled_skip_patterns.append(re.compile(p))
+            except re.error:
+                pass
+
     bindings: list[ResumeThumbBinding] = []
     orphan = 0
     excluded_indices: set[int] = set()
@@ -1068,14 +1116,14 @@ def _resume_thumb_bindings_and_orphans(
 
             if best_i is not None:
                 # PRD §B.1 C：距离检查——文字底部到卡片所属 speaker_band 顶部
-                # 超过阈值（50px baseline × scale_w）则不作为昵称
+                # 超过阈值（config.cross_band_distance_threshold_px × scale_w）则不作为昵称
                 _use_as_nickname = True
                 if band_idx is not None and band_idx >= 0:
                     _sb_list = getattr(layout_opt, "speaker_bands", ())
                     if band_idx < len(_sb_list):
                         _band_y0 = float(_sb_list[band_idx][0])
                         _gap = _band_y0 - best_bottom  # positive → block above band
-                        _threshold = 50.0 * float(scale_w)
+                        _threshold = float(config.cross_band_distance_threshold_px) * float(scale_w)
                         if _gap > _threshold:
                             _use_as_nickname = False
                 if _use_as_nickname:

{screenshot_vision_algorithm-0.3.8 → screenshot_vision_algorithm-0.3.10}/src/screenshot_vision_algorithm.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: screenshot_vision_algorithm
-Version: 0.3.8
+Version: 0.3.10
 Summary: Multi-platform screenshot vision algorithms: WeChat (card bbox, speaker bands, avatar detection, OCR helpers), XHS & Zhihu (planned)
 License: MIT
 Requires-Python: >=3.10