#WorkBuddy# 120 页"扫描件"版面还原：不去斜的 XY-cut 只有 0.64，校正 0.07° 斜角后冲到 0.92

import numpy as np
from PIL import Image

def deskew(ink):
    """投影方差法去斜：行墨迹方差最大的角度即校正角"""
    im = Image.fromarray(ink.astype(np.uint8) * 255)
    best_a, best_s = 0.0, -1
    for a in np.arange(-2.0, 2.01, 0.2):          # 粗搜
        r = np.array(im.rotate(a, resample=Image.BILINEAR, fillcolor=0))
        s = float(np.var(r.sum(1).astype(np.float64)))
        if s > best_s: best_s, best_a = s, a
    for a in np.arange(best_a - 0.2, best_a + 0.21, 0.05):  # 细搜
        r = np.array(im.rotate(a, resample=Image.BILINEAR, fillcolor=0))
        s = float(np.var(r.sum(1).astype(np.float64)))
        if s > best_s: best_s, best_a = s, a
    return np.array(im.rotate(best_a, resample=Image.BILINEAR, fillcolor=0)) > 120, float(best_a)

def detect_tables(ink):
    """>=2 条 380px 长横线且行距 <=60px → 锁为表格障碍物"""
    h, w = ink.shape
    line_rows = []
    for y in range(h):
        nz = np.nonzero(ink[y])[0]
        if len(nz) == 0: continue
        d = np.diff(nz); brk = np.nonzero(d > 1)[0]
        starts = np.concatenate([[0], brk + 1])
        ends = np.concatenate([brk, [len(nz) - 1]])
        run = max((nz[e] - nz[s] + 1) for s, e in zip(starts, ends))
        if run >= 380:
            line_rows.append(y)
    groups = []
    for y in line_rows:
        if groups and y - groups[-1][-1] <= 60: groups[-1].append(y)
        else: groups.append([y])
    boxes = []
    for g in groups:
        if len(g) < 2: continue
        xs = np.nonzero(ink[g[0]])[0]   # 首行横线横贯表宽 → 其起止即表格 x 范围
        boxes.append((int(xs[0]), g[0] - 4, int(xs[-1]), g[-1] + 2, "table"))
    return boxes

def xy_cut(ink, x0, y0, x1, y1, obstacles, out, min_hg=12, min_vg=26):
    """单刀递归 XY-cut：横竖一起比，每次只切最宽的一刀"""
    if x1 - x0 < 8 or y1 - y0 < 8: return
    sub = ink[y0:y1, x0:x1]
    if sub.sum() == 0: return
    for ob in obstacles:                       # 整块落在障碍物内 → 叶子
        if ob[0] <= x0 and x1 <= ob[2] and ob[1] <= y0 and y1 <= ob[3]:
            out.append({"bbox": [x0, y0, x1, y1], "type": ob[4]}); return
    def gaps(profile, ming):
        runs, st = [], None
        for i, v in enumerate(profile):
            if v == 0:
                st = i if st is None else st
            else:
                if st is not None and i - st >= ming: runs.append((st, i))
                st = None if v != 0 else st
        if st is not None and len(profile) - st >= ming: runs.append((st, len(profile)))
        return runs
    def ok(g, axis):                           # 切线不穿障碍物（用收缩 3px 的框判）
        for ox0, oy0, ox1, oy1, _ in obstacles:
            ox0, oy0, ox1, oy1 = ox0+3, oy0+4, ox1-3, oy1-4
            if axis == "x":
                gx0, gx1 = x0+g[0], x0+g[1]
                if gx1 > ox0 and gx0 < ox1 and oy1 > y0 and oy0 < y1 \
                   and ox0 >= x0-2 and ox1 <= x1+2 and oy0 >= y0-2 and oy1 <= y1+2:
                    return False
            else:
                gy0, gy1 = y0+g[0], y0+g[1]
                if gy1 > oy0 and gy0 < oy1 and ox1 > x0 and ox0 < x1 \
                   and ox0 >= x0-2 and ox1 <= x1+2 and oy0 >= y0-2 and oy1 <= y1+2:
                    return False
        return True
    xg = [g for g in gaps(sub.sum(0), min_vg) if ok(g, "x")]
    yg = [g for g in gaps(sub.sum(1), min_hg) if ok(g, "y")]
    best = None
    for axis, gs in (("x", xg), ("y", yg)):
        for g in gs:
            if best is None or g[1]-g[0] > best[0]: best = (g[1]-g[0], axis, g)
    if best is None:
        out.append({"bbox": [x0, y0, x1, y1], "type": "text"}); return
    _, axis, (a, b) = best
    if axis == "x":
        xy_cut(ink, x0, y0, x0+a, y1, obstacles, out, min_hg, min_vg)
        xy_cut(ink, x0+b, y0, x1, y1, obstacles, out, min_hg, min_vg)
    else:
        xy_cut(ink, x0, y0, x1, y0+a, obstacles, out, min_hg, min_vg)
        xy_cut(ink, x0, y0+b, x1, y1, obstacles, out, min_hg, min_vg)

请用 Python 完成文档图像版面还原，要求：
1. 输入 <图片目录>，输出 blocks.json（每块含 bbox、type、阅读序号）；
2. 流水线：投影方差法去斜(±2°粗搜0.2°+细搜0.05°) → 二值化 →
   长横线(>=380px、>=2条、行距<=60px)锁定表格 → 连通域(>=100x80)锁定插图 →
   单刀 XY-cut(水平垂直空隙一起比,每次切最宽一刀,切线不穿障碍物,行空隙阈值12px、列空隙阈值26px) →
   页眉页脚剔除(整块在页顶80px/页底60px内) → 类型判别(墨迹高度>28标题/18-28章节行/<=17且邻障碍物为图注) →
   阅读序(有第二栏时:跨轴块→左栏按y→右栏按y;单栏页全按y);
3. 评估：与真值比 IoU>=0.7 的切分 F1、IoU>=0.8 的块完整率、按真值序的逆序对率；
4. 所有中间结果落盘 JSON，报告每页耗时与去斜角误差。