screenshot-vision-algorithm 0.3.2__tar.gz → 0.3.8__tar.gz

{screenshot_vision_algorithm-0.3.2 → screenshot_vision_algorithm-0.3.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: screenshot_vision_algorithm
-Version: 0.3.2
+Version: 0.3.8
 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-0.3.2 → screenshot_vision_algorithm-0.3.8}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "screenshot_vision_algorithm"
-version = "0.3.2"
+version = "0.3.8"
 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-0.3.2 → screenshot_vision_algorithm-0.3.8}/src/screenshot_vision_algorithm/android/wechat/algorithms/badge_detection.py

@@ -218,7 +218,7 @@ def count_badges_per_row_roi(
 
     del kw  # reserved for forwards-compat
 
-    from collector.driver import (
+    from screenshot_vision_algorithm.android.wechat.algorithms.template_matching import (
         AVATAR_UNREAD_BADGE_LEFT_INSET_BASELINE,
         AVATAR_UNREAD_BADGE_RIGHT_PAD_BASELINE,
         CONVERSATION_AVATAR_RIGHT_EDGE_BASELINE,

{screenshot_vision_algorithm-0.3.2 → screenshot_vision_algorithm-0.3.8}/src/screenshot_vision_algorithm/android/wechat/algorithms/card_bbox.py

@@ -46,17 +46,6 @@ from typing import Iterable, Optional, Sequence
 import cv2
 import numpy as np
 
-#: Distance from the top of the ``favorite_label.png`` hit to the top of the
-#: enclosing note card, measured at 1080x2248 baseline. Used as an
-#: **upper-bound heuristic** for the first hit only (subsequent hits derive
-#: their top from the previous hit's bottom + ``FAVORITE_TAIL_OFFSET``).
-FAVORITE_TO_CARD_TOP_OFFSET_BASELINE = 421
-
-#: Gap from the bottom of the ``favorite_label.png`` hit to the card's
-#: actual bottom edge, at 1080x2248 baseline. Used to extend the card bbox
-#: past the label.
-FAVORITE_TAIL_OFFSET_BASELINE = 60
-
 #: ``favorite_label.png`` raster size @ 1080-wide baseline (`chat_profiles/README.md`).
 FAVORITE_LABEL_TEMPLATE_W_BASELINE = 70
 FAVORITE_LABEL_TEMPLATE_H_BASELINE = 60
@@ -166,7 +155,10 @@ def _compute_exact_card_and_bubble_bboxes(
     bgr_img: np.ndarray,
     screen_w: int,
     screen_h: int,
-) -> tuple[list[ThumbnailCard], list[BubbleBbox]]:
+) -> tuple[list[ThumbnailCard], list[BubbleBbox], bool]:
+    """Returns ``(cards, bubbles, pipeline_active)``; ``pipeline_active`` is
+    ``True`` when the pipeline found vlines / zones and actively decided which
+    anchors to keep or drop, ``False`` when it dropped out early (no texture)."""
     """PRD S6: vline-segment + hline-boundary card + bubble bbox detection.
 
     Steps:
@@ -248,7 +240,7 @@ def _compute_exact_card_and_bubble_bboxes(
     vd_lines = [vx for vx in vd_lines if vx - AR >= min_card_x_gap]
 
     if not vd_lines:
-        return [], []
+        return [], [], False
 
     # Pick best vline: max total mid pixels + longest continuous segment
     vline_scores: dict[int, float] = {}
@@ -274,7 +266,7 @@ def _compute_exact_card_and_bubble_bboxes(
             best_vx = vx
 
     if best_vx is None:
-        return [], []
+        return [], [], False
 
     # Step 3 — vline segments, computed per candidate vline. A fav-anchored
     # card may sit on a different right-edge vline than the frame-global best
@@ -306,7 +298,7 @@ def _compute_exact_card_and_bubble_bboxes(
     )
     zones_by_vx = {vx: _zones_for(vx) for vx in vline_order}
     if not any(zones_by_vx.values()):
-        return [], []
+        return [], [], False
 
     # Step 4 — card boundary hlines (span > 50% card_w, X-near vline,
     # max gray gradient >= 30 = white→gray boundary)
@@ -405,6 +397,7 @@ def _compute_exact_card_and_bubble_bboxes(
     gap_tol = max(1, int(round(GAP_HLINE_TO_NEXT_ZONE_MAX_BASELINE * sw)))
     cards: list[ThumbnailCard] = []
     prev_bottom: int | None = None  # anti-overlap: enforce card boundaries non-overlapping
+    any_anchor_seen: bool = False  # True if ≥1 anchor entered zone-matching loop
     for idx, hit in enumerate(ordered_hits):
         fav_cy = hit.y + hit.h // 2
 
@@ -413,6 +406,7 @@ def _compute_exact_card_and_bubble_bboxes(
             zones = zones_by_vx[vx]
             if not zones:
                 continue
+            any_anchor_seen = True  # anchor entered vline loop with non-empty zones
             card_hlines = hlines_by_vx[vx]
 
             def _snap(y1_: int, y2_: int) -> tuple[int, int]:
@@ -553,7 +547,7 @@ def _compute_exact_card_and_bubble_bboxes(
                 continue
             bubbles.append(BubbleBbox(top=by1, bottom=by2, left=AR, right=int(vx)))
 
-    return cards, bubbles
+    return cards, bubbles, any_anchor_seen
 
 
 def _compute_exact_card_bboxes(
@@ -561,13 +555,22 @@ def _compute_exact_card_bboxes(
     bgr_img: np.ndarray,
     screen_w: int,
     screen_h: int,
-) -> list[ThumbnailCard]:
+) -> tuple[list[ThumbnailCard], bool]:
     """Backward-compat wrapper — cards only (see
-    :func:`_compute_exact_card_and_bubble_bboxes`)."""
-    cards, _bubbles = _compute_exact_card_and_bubble_bboxes(
+    :func:`_compute_exact_card_and_bubble_bboxes`).
+
+    Returns:
+        (cards, pipeline_active):
+          - ``pipeline_active``: ``True`` when the pipeline found vlines
+            and zones (it *actively* decided to drop anchors without a
+            containing zone); ``False`` when the image had insufficient
+            texture (no vlines / no zones) and a legacy fallback is
+            appropriate.
+    """
+    cards, _bubbles, pipeline_active = _compute_exact_card_and_bubble_bboxes(
         ordered_hits, bgr_img, screen_w, screen_h,
     )
-    return cards
+    return cards, pipeline_active
 
 
 def derive_cards_and_bubbles(
@@ -591,7 +594,7 @@ def derive_cards_and_bubbles(
         return [], []
     ordered = sorted(hits, key=lambda h: h.y)
     try:
-        cards, bubbles = _compute_exact_card_and_bubble_bboxes(
+        cards, bubbles, _pipeline_active = _compute_exact_card_and_bubble_bboxes(
             ordered, chat_img, screen_w, screen_h,
         )
     except Exception:
@@ -622,15 +625,15 @@ def derive_cards(
               (we sort by y ourselves).
         screen_w / screen_h: the source screenshot's raw-image dimensions
                              (usually equal to ``device_info.screen_resolution``).
-        scale_w: ``screen_w / baseline_w`` (baseline = 1080). Used to scale
-                 ``FAVORITE_TO_CARD_TOP_OFFSET`` + ``FAVORITE_TAIL_OFFSET``
-                 so the logic survives across the 720 / 1080 / 1220 device
-                 whitelist.
-        chat_img: optional BGR image (as returned by ``cv2.imdecode``).
-                  When provided, uses the vline-segment + hline-boundary
-                  variance-based algorithm (PRD S6) for precise card bbox
-                  derivation. Falls back to legacy fixed-offset logic
-                  when ``chat_img`` is None.
+        scale_w: ``screen_w / baseline_w`` (baseline = 1080). Passed to
+                 :func:`drop_top_clamped_false_positive_cards` for
+                 false-positive filtering.
+        chat_img: required BGR image (as returned by ``cv2.imdecode``).
+                  The vline-segment + hline-boundary variance-based
+                  algorithm (PRD S6) is used for precise card bbox
+                  derivation. Returns ``[]`` when ``chat_img`` is
+                  ``None`` or the pipeline encounters an unrecoverable
+                  error.
 
     Returns:
         One :class:`ThumbnailCard` per hit, sorted top -> bottom. Empty list
@@ -642,61 +645,22 @@ def derive_cards(
 
     ordered = sorted(hits, key=lambda h: h.y)
 
-    # New algorithm: vline-segment + hline-boundary (PRD S6). When the
-    # pipeline ran successfully with non-empty cards its verdict is final —
-    # an anchor it dropped (no containing zone → "点对卡、记错框" guard)
-    # must NOT be resurrected by the legacy fixed-offset fallback. When the
-    # pipeline produces no cards (e.g. variance texture insufficient),
-    # legacy fixed-offset serves as a reasonable fallback.
-    if chat_img is not None and len(ordered) > 0:
+    # vline-segment + hline-boundary algorithm (PRD S6). Pipeline
+    # result is final — anchors it dropped (no containing zone → "点对卡、
+    # 记错框" guard) must NOT be resurrected by fallback logic.
+    if chat_img is not None:
         try:
-            exact_cards = _compute_exact_card_bboxes(
+            exact_cards, _pipeline_active = _compute_exact_card_bboxes(
                 ordered, chat_img, screen_w, screen_h,
             )
         except Exception:
-            pass
+            return []
         else:
-            if exact_cards:
-                return drop_top_clamped_false_positive_cards(
-                    exact_cards, scale_w=scale_w,
-                )
-            # exact_cards empty → fall through to legacy
-
-    # Legacy: fixed-offset algorithm
-    top_offset = int(round(FAVORITE_TO_CARD_TOP_OFFSET_BASELINE * scale_w))
-    tail_offset = int(round(FAVORITE_TAIL_OFFSET_BASELINE * scale_w))
-
-    cards: list[ThumbnailCard] = []
-    for idx, hit in enumerate(ordered):
-        if idx == 0:
-            card_top = max(0, hit.y1 - top_offset)
-        else:
-            prev_bottom = ordered[idx - 1].y2 + tail_offset
-            card_top = max(prev_bottom + 1, 0)
-
-        card_bottom = min(screen_h - 1, hit.y2 + tail_offset)
-
-        if card_bottom <= card_top:
-            continue
-
-        click_x = screen_w // 2
-        click_y = min(screen_h - 1, max(0, hit.y + hit.h // 2))
-        midline = screen_w / 2
-        click_side = "left" if click_x < midline else "right"
-
-        cards.append(ThumbnailCard(
-            index=idx,
-            top=card_top,
-            bottom=card_bottom,
-            left=0,
-            right=screen_w - 1,
-            click_x=click_x,
-            click_y=click_y,
-            click_side=click_side,
-            favorite_hit=hit,
-        ))
+            return drop_top_clamped_false_positive_cards(
+                exact_cards, scale_w=scale_w,
+            )
 
-    return drop_top_clamped_false_positive_cards(cards, scale_w=scale_w)
+    return []
 
 
 def drop_top_clamped_false_positive_cards(
@@ -874,12 +838,32 @@ def _card_height(card: ThumbnailCard) -> int:
     return card.bottom - card.top + 1
 
 
+#: 卡片高度匹配容差（像素）。同一张卡片在相邻两帧中 bbox 高度可能差 1-2px
+#: （back_to_chat 过渡动画导致列表微偏移），±5px 容差可覆盖此误差。
+_CARD_HEIGHT_TOLERANCE: int = 5
+
+
+def _heights_match(
+    a_heights: list[int], a_start: int,
+    b_heights: list[int], b_start: int,
+    length: int,
+) -> bool:
+    """Return True if a_heights[a_start:a_start+length] equals b_heights[b_start:b_start+length]
+    within ``_CARD_HEIGHT_TOLERANCE`` px per element."""
+    for k in range(length):
+        if abs(a_heights[a_start + k] - b_heights[b_start + k]) > _CARD_HEIGHT_TOLERANCE:
+            return False
+    return True
+
+
 def _longest_subsequence_match(
     a_heights: list[int],
     b_heights: list[int],
 ) -> tuple[int, int, int] | None:
     """Find the longest contiguous substring of a_heights that appears in b_heights.
 
+    Matching uses ± ``_CARD_HEIGHT_TOLERANCE`` px per element (PRD §10 第二步 容差)。
+
     Returns ``(a_start, b_start, length)`` or ``None``.
     If multiple matches share the same maximum length, picks the one with
     the smallest ``b_start`` (earliest in B).
@@ -894,9 +878,8 @@ def _longest_subsequence_match(
         for length in range(max_len, 0, -1):
             if best is not None and length < best[2]:
                 break
-            sub = a_heights[a_start:a_start + length]
             for b_start in range(len(b_heights) - length + 1):
-                if b_heights[b_start:b_start + length] == sub:
+                if _heights_match(a_heights, a_start, b_heights, b_start, length):
                     if (
                         best is None
                         or length > best[2]
@@ -971,8 +954,6 @@ def pick_next_unclicked_card(
 __all__ = [
     "FAVORITE_LABEL_TEMPLATE_W_BASELINE",
     "FAVORITE_LABEL_TEMPLATE_H_BASELINE",
-    "FAVORITE_TO_CARD_TOP_OFFSET_BASELINE",
-    "FAVORITE_TAIL_OFFSET_BASELINE",
     "FAV_LABEL_HLINE_EXCLUDE_PAD_BASELINE",
     "GAP_HLINE_TO_NEXT_ZONE_MAX_BASELINE",
     "PROCESSED_CARD_MATCH_MIN_OVERLAP_RATIO",

{screenshot_vision_algorithm-0.3.2 → screenshot_vision_algorithm-0.3.8}/src/screenshot_vision_algorithm/android/wechat/algorithms/speaker_band.py

@@ -26,7 +26,7 @@ from screenshot_vision_algorithm.android.wechat.algorithms.template_matching imp
 CHAT_BAND_FIRST_AVATAR_TOP_GAP_REJECT_BASELINE = 30
 
 #: 首个发言人筒上移量（一个头像高度，防止几何误差），见 PRD §5 补充2。
-CHAT_FIRST_BAND_TOP_EXTEND_BASELINE = 100
+CHAT_FIRST_BAND_TOP_EXTEND_BASELINE = 20
 
 #: 底部输入区预留（与主列表 scanner 一脉，baseline）。
 CHAT_COMPOSER_RESERVE_BOTTOM_BASELINE = 160
@@ -173,7 +173,7 @@ def detect_chat_side_avatar_ytops(
     y1 = max(0, bounds.y_top)
     y2 = min(h, bounds.y_bottom_excl)
     try:
-        from processor.left_avatar_column import (
+        from screenshot_vision_algorithm.android.wechat.algorithms.avatar_column import (
             detect_left_avatar_column_layout,
             left_avatar_ytops,
         )
@@ -254,7 +254,7 @@ def build_prd_speaker_vertical_bands(
         if not merged_starts or y - merged_starts[-1] >= min_sep:
             merged_starts.append(y)
 
-    # 首个筒上界上移（PRD §5 补充2）：上移 100px @1080，防几何误差与系统变化。
+    # 首个筒上界上移（PRD §5 补充2）：上移 20px @1080，防几何误差与系统变化。
     first_band_extend = int(round(int(first_band_top_extend_baseline) * sw))
 
     bands: list[tuple[int, int]] = []
@@ -275,6 +275,13 @@ def build_prd_speaker_vertical_bands(
         elif first_band_start < orphan_top[1]:
             orphan_top = (orphan_top[0], first_band_start)
 
+    # 无 orphan 时，首个 band 上界对齐 y_ct，保证纵向连续全覆盖
+    # （extend 较小且头像紧贴 y_ct 时可能产生小间隙）
+    if orphan_top is None and bands:
+        first_y0, first_y1 = bands[0]
+        if first_y0 > y_ct:
+            bands[0] = (y_ct, first_y1)
+
     return PrdChatVerticalLayout(
         content_bounds=bounds,
         orphan_top_band=orphan_top,

{screenshot_vision_algorithm-0.3.2 → screenshot_vision_algorithm-0.3.8}/src/screenshot_vision_algorithm/android/wechat/algorithms/template_matching.py

@@ -419,6 +419,43 @@ def detect_chat_title_more_dots(
     return more
 
 
+def detect_wechat_layer(
+    screen_bgr: np.ndarray,
+    scale_w: float,
+) -> str | None:
+    """4-layer classification — layout identical to the 162/162 3-device experiment.
+
+    Priority order (top to bottom):
+
+    * **L3**: ``detect_wechat_note_header`` hit
+    * **L2**: ``match_chat_back_chevron_core`` or ``match_chat_title_more_dots_core``
+      (no chrome / session-list-shell suppressor — the experiment used bare
+      ``_core`` hits + ``_CONTEXT_SUPPRESS`` post-filtering, and the
+      note_header gate is already exercised by the L3-takes-priority
+      ordering)
+    * **L1**: ``is_wechat_main_conversation_list_chrome`` (without pinned-row
+      requirement) or ``detect_wechat_main_bottom_tab_bar_four_columns``
+    * **None**: no layer recognised
+
+    Callers that also need external foreground-package gating (L0)
+    are responsible for checking the foreground app before calling
+    this function.
+    """
+    if detect_wechat_note_header(screen_bgr, scale_w) is not None:
+        return "L3"
+    if match_chat_back_chevron_core(screen_bgr, scale_w) is not None:
+        return "L2"
+    if match_chat_title_more_dots_core(screen_bgr, scale_w) is not None:
+        return "L2"
+    if is_wechat_main_conversation_list_chrome(
+        screen_bgr, scale_w, require_visible_pinned_row=False,
+    ):
+        return "L1"
+    if detect_wechat_main_bottom_tab_bar_four_columns(screen_bgr, scale_w):
+        return "L1"
+    return None
+
+
 def _chat_bottom_input_roi(screen_bgr: np.ndarray, scale_w: float) -> np.ndarray:
     """Bottom / keyboard-adjacent band for 话筒・表情・「+」（含输入框略抬高时的中区）。"""
     h, w = screen_bgr.shape[:2]
@@ -851,6 +888,12 @@ CONVERSATION_FIRST_ROW_TOP_BASELINE = 280
 #: Avatar column ends ~here @ 1080 baseline; nickname + preview middle band;
 # mirrors :meth:`collector.driver.TopGroupScanDriver.scan_pinned_groups`.
 CONVERSATION_AVATAR_RIGHT_EDGE_BASELINE = 220
+#: Horizontal inset left of ``CONVERSATION_AVATAR_RIGHT_EDGE_BASELINE`` for
+#: unread badge horizontal ROI (1080 baseline).
+AVATAR_UNREAD_BADGE_LEFT_INSET_BASELINE = 90
+#: Horizontal padding right of ``CONVERSATION_AVATAR_RIGHT_EDGE_BASELINE`` for
+#: unread badge horizontal ROI (1080 baseline).
+AVATAR_UNREAD_BADGE_RIGHT_PAD_BASELINE = 36
 CONVERSATION_RIGHT_SIDE_RESERVED_BASELINE = 220
 
 #: Horizontal ROI (1080 baseline) for list-side avatar centroid detection — left
@@ -1256,6 +1299,9 @@ def detect_wechat_main_bottom_tab_bar_four_columns(
     # Weaker tab may be 「我」单层笔画；最强一般为「通讯录」—— 拒绝单点极强、其余近空
     if fmin < fmax * 0.20:
         return False
+    # 任一列墨水>0.8%即认可（"最近"页面无"微信"栏会匹配）
+    if fmax < 0.008:
+        return False
     return True
 
 
@@ -1455,23 +1501,17 @@ def reposition_anchor_bottom_wechat_tap_lane_plausible(
     return True
 
 
-def _longest_true_run(mask: np.ndarray) -> int:
-    """最大连续 True 块的行数。"""
-    if not mask.any():
-        return 0
-    padded = np.concatenate(([False], mask, [False]))
-    edges = np.where(padded[:-1] != padded[1:])[0]
-    if len(edges) == 0:
-        return 0
-    runs = edges[1::2] - edges[::2]
-    return int(runs.max())
-
-
 def _recent_pull_top_heading_likely(
     screen_bgr: np.ndarray,
     scale_w: float,
 ) -> bool:
-    """顶区「最近」类标题：亮底深字 **或** 暗底浅字，竖向窄中带。"""
+    """顶区「最近」类标题：亮底深字 **或** 暗底浅字，竖向窄中带。
+
+    - 优先检测 dark-pixel 墨迹密度 + 竖向跨度 span
+    - 再尝试 gray>96 亮像素（小米/华为 ROM 差异化兼容）
+    - Honor／暗色主题下 ink_d 异常高，走 gray>150 亮字簇路径，
+      用 **最大连续亮行 run** 排除散点噪声 vs 真标题集群。
+    """
     h, w = screen_bgr.shape[:2]
     sw = max(scale_w, 1e-6)
     y1 = int(round(222 * sw))
@@ -1492,27 +1532,72 @@ def _recent_pull_top_heading_likely(
 
     _, mask_dark = cv2.threshold(gray, 80, 255, cv2.THRESH_BINARY_INV)
     ink_d = float(cv2.countNonZero(mask_dark)) / area
+
+    #  路径 1：暗墨标题（D1/D2 浅色主题首选）
     if 0.014 <= ink_d <= 0.48:
         row_dark = (mask_dark.sum(axis=1) // 255).astype(np.int32)
         has_row = row_dark > 9
         if has_row.any():
-            # 最大连续块跨度：低分辨率下散点不会拉长 span（2026-06-13）
-            block_span = _longest_true_run(has_row)
-            if block_span <= int(round(70 * sw)):
+            idx = np.where(has_row)[0]
+            span_rows = int(idx.max()) - int(idx.min()) + 1
+            if span_rows <= int(round(70 * sw)):
                 return True
 
+    #  路径 2：暗色主题下 ink_d 异常高 → 走亮字簇（D3 Honor 暗色主题）
+    if ink_d > 0.48:
+        bright = (gray > 150).astype(np.uint8)
+        ink_b = float(np.sum(bright)) / area
+        if 0.004 <= ink_b <= 0.10:
+            row_counts = bright.sum(axis=1).astype(np.int32)
+            max_run = 0
+            cur_run = 0
+            for rc in row_counts:
+                if rc > 3:
+                    cur_run += 1
+                    if cur_run > max_run:
+                        max_run = cur_run
+                else:
+                    cur_run = 0
+            if 8 <= max_run <= int(round(80 * sw)):
+                return True
+
+    #  路径 3：不依赖字体颜色的通用亮底（小米 / 华为 ROM 兼容）
     light = (gray > 96).astype(np.uint8)
     ink_l = float(np.sum(light)) / area
     if 0.006 <= ink_l <= 0.32:
         row_lit = light.sum(axis=1).astype(np.int32)
         has_row = row_lit > 5
         if has_row.any():
-            block_span = _longest_true_run(has_row)
-            if block_span <= int(round(72 * sw)):
+            idx = np.where(has_row)[0]
+            span_rows = int(idx.max()) - int(idx.min()) + 1
+            if span_rows <= int(round(72 * sw)):
                 return True
     return False
 
 
+def is_wechat_recent_page_chrome(
+    screen_bgr: np.ndarray,
+    scale_w: float,
+) -> bool:
+    """Detect WeChat 「最近」小程序抽屉页。
+
+    Two conditions must BOTH hold:
+    1. Top mid-band contains a compact heading-like dark/bright ink cluster
+       (:func:`_recent_pull_top_heading_likely`).
+    2. The WeChat 4-tab bottom bar is **absent** — the 「最近」drawer
+       replaces it with a dark/empty strip.
+
+    Condition (2) prevents false positives on the normal conversation list
+    where group names and message previews can produce heading-like ink
+    densities in the mid-band.
+    """
+    if not _recent_pull_top_heading_likely(screen_bgr, scale_w):
+        return False
+    if detect_wechat_main_bottom_tab_bar_four_columns(screen_bgr, scale_w):
+        return False
+    return True
+
+
 def is_wechat_recent_pull_anchor_early_stop_frame(
     screen_bgr: np.ndarray,
     scale_w: float,
@@ -1520,15 +1605,12 @@ def is_wechat_recent_pull_anchor_early_stop_frame(
     tab_x: int,
     tab_y: int,
 ) -> bool:
-    """§步骤1(2) 早停：顶区「最近」类标题 **且** 与点击同标的底栏邻域 ``(tab_x,tab_y)`` 可点条带。
+    """§步骤1(2) 早停：heading 命中 + 底栏四列缺失。
 
-    底栏侧与归位 ``adb.tap(tab_x, tab_y)`` 共用 :func:`reposition_anchor_bottom_wechat_tap_lane_plausible`，
-    **不**再单独用四 Tab / 胶囊中区视检。"""
-    if not reposition_anchor_bottom_wechat_tap_lane_plausible(
-            screen_bgr, scale_w, tab_x=tab_x, tab_y=tab_y,
-    ):
-        return False
-    return _recent_pull_top_heading_likely(screen_bgr, scale_w)
+    通过 :func:`is_wechat_recent_page_chrome` 双条件把关 —
+    heading 启发式 + 四列缺失保证不会在普通会话列表上误触发。
+    ``tab_x`` / ``tab_y`` 保留为兼容参数，内部不参与判断。"""
+    return is_wechat_recent_page_chrome(screen_bgr, scale_w)
 
 
 def split_conversation_rows(
@@ -2133,6 +2215,8 @@ __all__ = [
     "LIST_AVATAR_BADGE_ORANGE_AUX_VAL_MIN",
     "CONVERSATION_ROW_HEIGHT_BASELINE",
     "CONVERSATION_AVATAR_RIGHT_EDGE_BASELINE",
+    "AVATAR_UNREAD_BADGE_LEFT_INSET_BASELINE",
+    "AVATAR_UNREAD_BADGE_RIGHT_PAD_BASELINE",
     "CONVERSATION_LIST_AVATAR_ROI_LEFT_BASELINE",
     "CONVERSATION_LIST_AVATAR_ROI_RIGHT_BASELINE",
     "CONVERSATION_LIST_AVATAR_MEDIAN_X_HALF_WIDTH_BASELINE",

{screenshot_vision_algorithm-0.3.2 → screenshot_vision_algorithm-0.3.8}/src/screenshot_vision_algorithm/android/wechat/ocr/avatar_guard.py

@@ -19,7 +19,7 @@ except ImportError:
     cv2 = None  # type: ignore[assignment]
 
 try:
-    from processor.left_avatar_column import AvatarCentroid
+    from screenshot_vision_algorithm.android.wechat.algorithms.avatar_column import AvatarCentroid
 except ImportError:
     AvatarCentroid = None  # type: ignore[misc, assignment]
 

{screenshot_vision_algorithm-0.3.2 → screenshot_vision_algorithm-0.3.8}/src/screenshot_vision_algorithm/android/wechat/ocr/nickname_binding.py

@@ -351,6 +351,9 @@ _PURE_TIME_RE = re.compile(r"^\d{1,2}:\d{2}(:\d{2})?$")
 _WEEKDAY_TIME_RE = re.compile(
     r"^周[一二三四五六日天].*?(?:上午|下午|晚上|凌晨)?\s*\d{0,2}[:：]?\d{0,2}"
 )
+_YESTERDAY_TODAY_TIME_RE = re.compile(
+    r"^(?:昨天|今天)\s*\d{0,2}:?\d{0,2}$"
+)
 _MOBILE_PHONE_SUFFIX_RE = re.compile(r"1\d{10}\s*$")
 
 
@@ -388,6 +391,8 @@ def _is_chat_timestamp_or_divider(text: str) -> bool:
         return True
     if _WEEKDAY_TIME_RE.match(t):
         return True
+    if _YESTERDAY_TODAY_TIME_RE.match(t):
+        return True
     return False
 
 
@@ -411,7 +416,7 @@ def _is_avatar_column_nickname_row(
     avatar_layout: Any = None,
 ) -> bool:
     try:
-        from processor.left_avatar_column import nickname_bbox_in_avatar_column
+        from screenshot_vision_algorithm.android.wechat.algorithms.avatar_column import nickname_bbox_in_avatar_column
     except ImportError:
         nickname_bbox_in_avatar_column = None  # type: ignore[assignment,misc]
     if avatar_layout is not None and nickname_bbox_in_avatar_column is not None:
@@ -446,7 +451,7 @@ def _avatar_anchor_for_block(
     if avatar_layout is None or getattr(avatar_layout, "empty", True):
         return None
     try:
-        from processor.left_avatar_column import find_avatar_anchor_for_nickname_bbox
+        from screenshot_vision_algorithm.android.wechat.algorithms.avatar_column import find_avatar_anchor_for_nickname_bbox
     except ImportError:
         return None
     return find_avatar_anchor_for_nickname_bbox(
@@ -463,9 +468,9 @@ def _load_left_avatar_layout(
     if png_path is None:
         return None
     try:
-        from processor.chat_bubble_geometry_prd import compute_chat_content_vertical_bounds
-        from processor.left_avatar_column import detect_left_avatar_column_layout
-        from processor.nickname_avatar_guard import load_png_bgr
+        from screenshot_vision_algorithm.android.wechat.algorithms.speaker_band import compute_chat_content_vertical_bounds
+        from screenshot_vision_algorithm.android.wechat.algorithms.avatar_column import detect_left_avatar_column_layout
+        from screenshot_vision_algorithm.android.wechat.ocr.avatar_guard import load_png_bgr
     except ImportError:
         return None
     arr = load_png_bgr(str(png_path))
@@ -595,7 +600,7 @@ def _nickname_passes_avatar_guard(
     if bgr_guard is None:
         return not config.require_avatar_guard_for_nickname
     try:
-        from processor.nickname_avatar_guard import avatar_roi_pass
+        from screenshot_vision_algorithm.android.wechat.ocr.avatar_guard import avatar_roi_pass
     except ImportError:
         return not config.require_avatar_guard_for_nickname
 
@@ -687,13 +692,13 @@ def _load_prd_layout_optional(
     if png_path is None or not config.use_prd_chat_vertical_bands:
         return None
     try:
-        from processor.chat_bubble_geometry_prd import (
+        from screenshot_vision_algorithm.android.wechat.algorithms.speaker_band import (
             build_prd_speaker_vertical_bands,
             compute_chat_content_vertical_bounds,
             detect_chat_side_avatar_ytops,
             merge_avatar_ytops_time_order,
         )
-        from processor.nickname_avatar_guard import load_png_bgr
+        from screenshot_vision_algorithm.android.wechat.ocr.avatar_guard import load_png_bgr
     except ImportError:
         return None
     arr = load_png_bgr(str(png_path))
@@ -962,7 +967,17 @@ def _resume_thumb_bindings_and_orphans(
             _cy1, _cy2 = float(tbb[1]), float(tbb[3])
             for _bi, (_b0, _b1) in enumerate(_sb_list):
                 if _cy2 > float(_b0) and _cy1 < float(_b1):
-                    band_idx = _bi  # 相交 → 候选；持续覆盖 → 最后一个（下方）胜出
+                        band_idx = _bi  # 相交 → 候选；持续覆盖 → 最后一个（下方）胜出
+
+        # ── 如果卡片与 orphan_top_band 相交 → 直接记为找不到归属 ──
+        if layout_opt is not None:
+            ot = getattr(layout_opt, "orphan_top_band", None)
+            if ot is not None:
+                oy0, oy1 = float(ot[0]), float(ot[1])
+                _cy1, _cy2 = float(tbb[1]), float(tbb[3])
+                if _cy2 > oy0 and _cy1 < oy1:
+                    orphan += 1
+                    continue
 
         matched_nickname: Optional[str] = None
 
@@ -1177,7 +1192,7 @@ def first_attrib_verbatim_display_line_and_bbox(
     scale_w = float(original_width) / 1080.0
     if png_path is not None:
         try:
-            from processor.nickname_avatar_guard import load_png_bgr
+            from screenshot_vision_algorithm.android.wechat.ocr.avatar_guard import load_png_bgr
 
             bgr_guard = load_png_bgr(str(png_path))
         except ImportError:
@@ -1748,7 +1763,7 @@ def extract_nicknames(
     bgr_guard = None
     if png_path is not None:
         try:
-            from processor.nickname_avatar_guard import load_png_bgr
+            from screenshot_vision_algorithm.android.wechat.ocr.avatar_guard import load_png_bgr
 
             bgr_guard = load_png_bgr(str(png_path))
         except ImportError:

{screenshot_vision_algorithm-0.3.2 → screenshot_vision_algorithm-0.3.8}/src/screenshot_vision_algorithm.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: screenshot_vision_algorithm
-Version: 0.3.2
+Version: 0.3.8
 Summary: Multi-platform screenshot vision algorithms: WeChat (card bbox, speaker bands, avatar detection, OCR helpers), XHS & Zhihu (planned)
 License: MIT
 Requires-Python: >=3.10