Files
agent/skills/document-photo-to-pdf/doc_deskew/compose.py
T
kicer e39273e97f feat(document-photo-to-pdf): raw PNG 页图导出到 out/,PDF 从 raw 派生
出件改为一次透视:PIL 按四边形自然尺寸 warp 出 raw 页图(只裁剪+拉正,不调色、不缩放、无白边),直接写 out/<PDF 主名>-NN.png;PDF 页再由 raw 调色 + LANCZOS 缩到画布内容区派生。相比「PDF 目标尺寸一次 + raw 自然尺寸一次」少一次 warp,10MP 自然尺寸从 numpy 的 3.0s/页降到 PIL 的 0.3s/页。

不新增输入参数:每次运行 PDF 与 PNG 全出,AI 不需要判断格式;SKILL.md 只补 out/ 说明、images 返回字段,以及 description 里的产出(PDF + out/ 下的 PNG 原图)。

页序先定后渲染,PNG 编号跟最终 PDF 页序;PDF 落盘成功后才清理同前缀的多余编号(失败保留上一份),不碰 out/ 里其它文件。numpy warp 与旧 warp_page 保留作回退路径。

真实样张(task-20261021-004,3 页)A/B:state.json 完全一致;PDF 逐页平均差 0.7 / 0.7 / 1.2 灰度级;raw PNG 与 numpy 自然尺寸 warp 平均差 0.02;3 页端到端 4.8s,PNG 0.58/0.69/0.90MB。注:doc_deskew/smoke_test.py 自 fe1c6d1 起未同步返回契约,属既有问题,本次未改。
2026-10-24 09:54:37 +08:00

225 lines
8.1 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# -*- coding: utf-8 -*-
"""出页与合成:批次方向、目标尺寸、页面拼版、PDF 保存。
坐标与尺寸约定:
- 所有 quad 都是“转正后原图”坐标;
- 目标尺寸把 PDF 画布内容区(页面尺寸 - 四周留白)的一条边填满,可放大低分辨率照片,
避免同一批页面因原图像素多少而大小不一;
- raw 页图(裁剪 + 透视拉正、自然尺寸、不调色)由 warp_raw 产出,PDF 页从它派生。
"""
import os
import statistics
import numpy as np
from PIL import Image
from homography import brighten, warp_perspective, warp_perspective_pil
# ---------------------------------------------------------------- 纸张
def resolve_paper(paper, P):
"""把 paper 参数解析成 {name, canvas_mm, ratio}。
auto 不指定标准纸面比例;画布用 A4,比例由调用方按批次中位数决定
a4/letter… 画布和纸面比例都用该标准纸
数字 画布用 A4,纸面比例用该数字(宽/高)
"""
r = P["render"]
name = str(paper or r["default_paper"]).strip().lower()
a4 = r["papers_mm"]["a4"]
if name in ("", "auto"):
return {"name": "auto", "canvas_mm": a4, "ratio": None}
if name in r["papers_mm"]:
w_mm, h_mm = r["papers_mm"][name]
return {"name": name, "canvas_mm": (w_mm, h_mm),
"ratio": float(w_mm) / float(h_mm)}
try:
ratio = float(name)
except ValueError:
raise ValueError("未知纸张:%s(可选 %s / 数字宽高比 / auto)"
% (paper, "/".join(r["papers_mm"])))
if not (0.02 < ratio < 50.0):
raise ValueError("纸张比例超出合理范围:%s" % paper)
return {"name": "custom", "canvas_mm": a4, "ratio": ratio}
def canvas_px(paper, orientation, P):
"""按 dpi 换算画布像素;orientation=landscape 时交换宽高。"""
w_mm, h_mm = paper["canvas_mm"]
dpi = float(P["render"]["dpi"])
cw = int(round(w_mm / 25.4 * dpi))
ch = int(round(h_mm / 25.4 * dpi))
if orientation == "landscape":
cw, ch = ch, cw
return max(1, cw), max(1, ch)
# ---------------------------------------------------------------- 方向
def aspect_of(quad):
"""四边形在屏幕坐标下的自然宽高(像素)。"""
q = np.asarray(quad, dtype=np.float64)
w = (np.linalg.norm(q[1] - q[0]) + np.linalg.norm(q[2] - q[3])) / 2.0
h = (np.linalg.norm(q[3] - q[0]) + np.linalg.norm(q[2] - q[1])) / 2.0
return max(float(w), 1e-6), max(float(h), 1e-6)
def classify(aspect, deadband):
if aspect > 1.0 + deadband:
return "landscape"
if aspect < 1.0 - deadband:
return "portrait"
return "square"
def decide_orientation(aspects, mode, P):
"""返回 (batch_orientation, source, median_aspect)。"""
dead = P["gate"]["aspect_deadband"]
median = float(statistics.median(aspects)) if aspects else 1.0
if mode == "none":
return "mixed", "none", median
if mode in ("portrait", "landscape"):
return mode, "forced", median
classes = [classify(a, dead) for a in aspects]
n_land = sum(1 for c in classes if c == "landscape")
n_port = sum(1 for c in classes if c == "portrait")
if n_land > n_port:
return "landscape", "majority", median
if n_port > n_land:
return "portrait", "majority", median
if n_land == 0 and n_port == 0:
return "portrait", "default", median
return "portrait", "tie", median
def rotate_quad_cw(quad):
"""把角点序列旋转一位,等价于把页面顺时针转 90°。"""
q = list(quad)
return [q[3], q[0], q[1], q[2]]
# ---------------------------------------------------------------- 目标尺寸
def target_size(quad, paper, page_orientation, P, out_ratio=None):
"""返回 (target_w, target_h, (canvas_w, canvas_h))。
out_ratio 是 auto 模式下按批次中位比例算出的输出宽高比;None 表示按本页
自然比例。两种分支都把内容区的一条边填满:低分辨率照片会被放大,从而
避免同一批页面因原图尺寸/分辨率不同而大小不一。
"""
w_src, h_src = aspect_of(quad)
cw, ch = canvas_px(paper, page_orientation, P)
margin = int(P["render"]["margin_px"])
cont_w = max(1, cw - 2 * margin)
cont_h = max(1, ch - 2 * margin)
if paper["ratio"] is not None:
# 强制纸面比例:宽高比纠正透视压缩。
r = (paper["ratio"] if page_orientation == "portrait"
else 1.0 / paper["ratio"])
else:
r = out_ratio
if r is not None:
tw = float(cont_w)
th = tw / r
if th > cont_h:
th = float(cont_h)
tw = th * r
else:
# auto 且未形成批次比例:按四边形自然比例填满内容区(可放大)。
s = min(float(cont_w) / w_src, float(cont_h) / h_src)
tw, th = w_src * s, h_src * s
return max(1, int(round(tw))), max(1, int(round(th))), (cw, ch)
def warp_raw(img, quad):
"""只做裁剪 + 透视拉正:自然尺寸、不调色。raw 页图的唯一来源。"""
return warp_perspective_pil(img, quad, None)
def page_for_pdf(raw, size, P):
"""从 raw 页图派生 PDF 页:先调色,再按目标尺寸缩放(LANCZOS)。
与旧 warp_page 的差别只在采样顺序:先在自然尺寸做透视(超采样),
再缩到画布内容区,缩小时比直接按目标尺寸采样更少混叠。
"""
r = P["render"]
spec = P["enhance"].get(r.get("enhance") or r["default_enhance"])
page = tone(raw, spec, r["brighten_target"], r["brighten_pct"])
if page.size != tuple(size):
page = page.resize(tuple(size), Image.LANCZOS)
return page
def warp_page(img, quad, size, P):
"""旧路径:直接按目标尺寸 warp + 调色。PIL 实现出问题时回退用。"""
r = P["render"]
page = warp_perspective(img, quad, size,
chunk_rows=r["warp_chunk_rows"])
spec = P["enhance"].get(r.get("enhance") or r["default_enhance"])
return tone(page, spec, r["brighten_target"], r["brighten_pct"])
def tone(img, spec, target=245, pct=97):
"""按调色档做整体映射;spec=None 表示原样返回。
只做“分位数 -> 目标亮度”的等比缩放(加可选灰度),不碰像素间的
局部关系,所以同一张图重跑结果一致,也不会重绘文字边缘。
"""
if not spec:
return img
if spec.get("gray"):
img = img.convert("L").convert("RGB")
return brighten(img, spec.get("target", target), spec.get("pct", pct))
# ---------------------------------------------------------------- PDF
def build_pdf(pages, out_pdf, canvases, P):
"""把已校正的页面图居中放进各自画布,按顺序写 PDF。
pages / canvases 一一对应;画布可能因 orientation=none 而不同。
"""
r = P["render"]
margin = int(r["margin_px"])
dpi = int(r["dpi"])
composed = []
meta = []
for img, (cw, ch) in zip(pages, canvases):
cw, ch = max(1, int(cw)), max(1, int(ch))
cont_w = max(1, cw - 2 * margin)
cont_h = max(1, ch - 2 * margin)
s = min(1.0, float(cont_w) / img.width, float(cont_h) / img.height)
if s < 1.0:
nw = max(1, int(round(img.width * s)))
nh = max(1, int(round(img.height * s)))
img = img.resize((nw, nh), Image.LANCZOS)
else:
nw, nh = img.width, img.height
canvas = Image.new("RGB", (cw, ch), "white")
canvas.paste(img, ((cw - nw) // 2, (ch - nh) // 2))
composed.append(canvas)
meta.append({"canvas": [cw, ch], "pasted": [nw, nh]})
if not composed:
raise ValueError("没有可合成的页面")
out_dir = os.path.dirname(os.path.abspath(out_pdf))
os.makedirs(out_dir, exist_ok=True)
tmp = out_pdf + ".part.pdf"
try:
composed[0].save(tmp, "PDF", resolution=dpi, save_all=True,
append_images=composed[1:])
os.replace(tmp, out_pdf)
finally:
if os.path.exists(tmp):
try:
os.remove(tmp)
except OSError:
pass
return meta