自动识别边框

This commit is contained in:
2026-10-22 22:11:33 +08:00
parent 61245b1c65
commit deb622095f
6 changed files with 682 additions and 33 deletions
@@ -0,0 +1,500 @@
# -*- coding: utf-8 -*-
"""边线检测 + 纸角候选提取(从 edge_clean_c.py v3 抽取的精简模块)。
对外只暴露一个函数:
detect(image_path, max_side=1600) -> (corners, info)
corners:原图像素坐标系下的四角 [(x, y)],顺序 左上、右上、右下、左下;
info :附加信息(处理尺寸、缩放比、各边覆盖率、边线掩膜),供出图与诊断用。
注意:本模块只产出**候选**坐标。它不写 corners.json,也不参与出件判定——
最终角点必须以人工读角标图的结果为准。
"""
import numpy as np
from PIL import Image, ImageOps, ImageFilter
from collections import deque
# ---------------- 基础工具 ----------------
def load_and_scale(path, max_side=1600):
pil = ImageOps.exif_transpose(Image.open(path)).convert("RGB")
w, h = pil.size
s = max_side / max(w, h)
if s < 1:
pil = pil.resize((int(round(w * s)), int(round(h * s))), Image.LANCZOS)
return pil, np.asarray(pil).astype(np.float32) / 255.0
def rgb_to_sv(arr):
maxc = arr.max(axis=-1)
minc = arr.min(axis=-1)
delta = maxc - minc
S = np.zeros_like(maxc, dtype=np.float32)
np.divide(delta, maxc, out=S, where=maxc > 1e-6)
return S.astype(np.float32), maxc.astype(np.float32)
def luminance(arr):
return (0.299 * arr[..., 0] + 0.587 * arr[..., 1] + 0.114 * arr[..., 2]).astype(np.float32)
def box_mean(img, r):
"""O(1)/像素的方框均值(积分图),r 为半径。"""
img = np.asarray(img, dtype=np.float32)
H, W = img.shape
pad = np.zeros((H + 1, W + 1), dtype=np.float32)
pad[1:, 1:] = img
ii = pad.cumsum(0).cumsum(1)
ys = np.arange(H)
xs = np.arange(W)
y0 = np.clip(ys - r, 0, H)
y1 = np.clip(ys + r + 1, 0, H)
x0 = np.clip(xs - r, 0, W)
x1 = np.clip(xs + r + 1, 0, W)
S = (ii[np.ix_(y1, x1)] - ii[np.ix_(y0, x1)]
- ii[np.ix_(y1, x0)] + ii[np.ix_(y0, x0)])
cnt = (y1 - y0)[:, None] * (x1 - x0)[None, :]
return S / np.maximum(cnt, 1).astype(np.float32)
def gaussian_blur_np(img, sigma):
radius = int(np.ceil(sigma * 3))
if radius < 1:
return img.astype(np.float32)
x = np.arange(-radius, radius + 1, dtype=np.float32)
k = np.exp(-x * x / (2 * sigma * sigma))
k /= k.sum()
H, W = img.shape
p = np.pad(img, radius, mode="edge")
tmp = np.zeros((H, W), dtype=np.float32)
for i, kv in enumerate(k):
tmp += kv * p[radius:radius + H, i:i + W]
p = np.pad(tmp, ((radius, radius), (0, 0)), mode="edge")
out = np.zeros((H, W), dtype=np.float32)
for i, kv in enumerate(k):
out += kv * p[i:i + H, :]
return out
def convolve3x3(img, kernel):
H, W = img.shape
p = np.pad(img, 1, mode="edge")
out = np.zeros((H, W), dtype=np.float32)
for i in range(3):
for j in range(3):
kv = kernel[i, j]
if kv != 0:
out += kv * p[i:i + H, j:j + W]
return out
def sobel_magnitude(gray):
kx = np.array([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], dtype=np.float32)
ky = np.array([[-1, -2, -1], [0, 0, 0], [1, 2, 1]], dtype=np.float32)
gx = convolve3x3(gray, kx)
gy = convolve3x3(gray, ky)
return np.sqrt(gx * gx + gy * gy)
# ---------------- 形态学(Pillow Max/MinFilter)----------------
def _odd(size):
size = int(size)
if size < 3:
size = 3
if size % 2 == 0:
size += 1
return size
def dilate(mask, size):
pil = Image.fromarray((mask.astype(np.uint8) * 255))
return np.asarray(pil.filter(ImageFilter.MaxFilter(_odd(size)))) > 127
def erode(mask, size):
pil = Image.fromarray((mask.astype(np.uint8) * 255))
return np.asarray(pil.filter(ImageFilter.MinFilter(_odd(size)))) > 127
def closing(mask, size):
return erode(dilate(mask, size), size)
def opening(mask, size):
return dilate(erode(mask, size), size)
# ---------------- 连通域 & 填孔(降采样图上做)----------------
def downsample_any(mask, factor):
if factor <= 1:
return mask
H, W = mask.shape
H2, W2 = H // factor, W // factor
return mask[:H2 * factor, :W2 * factor].reshape(
H2, factor, W2, factor).any(axis=(1, 3))
def upsample_nearest(mask_small, target_h, target_w):
h, w = mask_small.shape
yi = np.linspace(0, h - 1, target_h).astype(np.int32)
xi = np.linspace(0, w - 1, target_w).astype(np.int32)
return mask_small[np.ix_(yi, xi)]
def fill_holes(mask_small):
H, W = mask_small.shape
bg = ~mask_small
visited = np.zeros((H, W), dtype=bool)
q = deque()
for x in range(W):
if bg[0, x] and not visited[0, x]:
visited[0, x] = True
q.append((0, x))
if bg[H - 1, x] and not visited[H - 1, x]:
visited[H - 1, x] = True
q.append((H - 1, x))
for y in range(H):
if bg[y, 0] and not visited[y, 0]:
visited[y, 0] = True
q.append((y, 0))
if bg[y, W - 1] and not visited[y, W - 1]:
visited[y, W - 1] = True
q.append((y, W - 1))
while q:
y, x = q.popleft()
if y > 0 and bg[y - 1, x] and not visited[y - 1, x]:
visited[y - 1, x] = True
q.append((y - 1, x))
if y < H - 1 and bg[y + 1, x] and not visited[y + 1, x]:
visited[y + 1, x] = True
q.append((y + 1, x))
if x > 0 and bg[y, x - 1] and not visited[y, x - 1]:
visited[y, x - 1] = True
q.append((y, x - 1))
if x < W - 1 and bg[y, x + 1] and not visited[y, x + 1]:
visited[y, x + 1] = True
q.append((y, x + 1))
return mask_small | (bg & ~visited)
def largest_connected(mask_small):
H, W = mask_small.shape
visited = np.zeros((H, W), dtype=bool)
best = np.zeros_like(mask_small)
best_n = 0
ys, xs = np.where(mask_small)
for y0, x0 in zip(ys, xs):
if visited[y0, x0]:
continue
stack = [(y0, x0)]
visited[y0, x0] = True
comp = []
while stack:
y, x = stack.pop()
comp.append((y, x))
for dy, dx in ((-1, -1), (-1, 0), (-1, 1), (0, -1),
(0, 1), (1, -1), (1, 0), (1, 1)):
ny, nx = y + dy, x + dx
if 0 <= ny < H and 0 <= nx < W and mask_small[ny, nx] and not visited[ny, nx]:
visited[ny, nx] = True
stack.append((ny, nx))
if len(comp) > best_n:
best_n = len(comp)
best = np.zeros_like(mask_small)
for y, x in comp:
best[y, x] = True
return best
# ---------------- 纸面掩膜(整形后只用于限定搜索区域)----------------
def otsu_threshold(score):
"""一维 Otsu:返回使类间方差最大的阈值"""
hist, bin_edges = np.histogram(score, bins=256, range=(0.0, 1.0))
hist = hist.astype(np.float64)
total = hist.sum()
if total <= 0:
return 0.5
p = hist / total
centers = (bin_edges[:-1] + bin_edges[1:]) * 0.5
w0 = np.cumsum(p)
w1 = 1.0 - w0
m0 = np.cumsum(p * centers)
mt = m0[-1]
with np.errstate(divide="ignore", invalid="ignore"):
mu0 = m0 / np.maximum(w0, 1e-12)
mu1 = (mt - m0) / np.maximum(w1, 1e-12)
between = w0 * w1 * (mu0 - mu1) ** 2
between = np.nan_to_num(between, nan=-1.0)
return float(centers[int(np.argmax(between))])
def paper_mask(arr, debug=False):
H, W = arr.shape[:2]
S, V = rgb_to_sv(arr)
# 纸面近白:亮度高、饱和低;木桌面偏色暗。用 Otsu 自动定阈值,
# 再叠加硬约束 V>0.50 防止桌面亮反射被并入。
score = (V * (1.0 - S)).astype(np.float32)
th = otsu_threshold(score)
th = max(th, 0.50)
bright = (score > th) & (V > 0.50) & (S < 0.45)
if debug:
print(f"[paper_mask] otsu={th:.3f} bright frac={bright.mean():.3f}")
factor = max(1, min(H, W) // 400)
bs = downsample_any(bright, factor)
size = _odd(max(5, min(bs.shape) // 60))
m = closing(bs, size)
m = fill_holes(m)
m = largest_connected(m)
m = closing(m, _odd(max(5, min(bs.shape) // 80))) # 抹掉内凹缺口/阶梯
m = fill_holes(m)
# 掩膜轻微内收:纸边最外侧 1~2px 常被阴影/暗边吃掉,内收后环带更贴合真实纸边
m = erode(m, 3)
paper = upsample_nearest(m, H, W) if factor > 1 else m
if debug:
Image.fromarray((paper * 255).astype(np.uint8)).save("dbg_v3_paper.png")
return paper
def boundary_band(paper, k):
"""纸面掩膜边界两侧各 k 像素的环带。"""
k = _odd(k)
return dilate(paper, k) & ~erode(paper, k)
# ---------------- 边缘图(局部自适应阈值)----------------
def edge_map(arr, local_gain=1.55, low=0.035):
H = arr.shape[0]
gray = luminance(arr)
sigma = max(1.0, H / 800.0)
blurred = gaussian_blur_np(gray, sigma)
mag = sobel_magnitude(blurred)
local = box_mean(mag, max(3, H // 50))
th = np.maximum(low, local * local_gain)
return mag > th
def boundary_step_filter(arr, paper, r_small=10, r_big=24, step_th=0.055):
"""跨纸边亮度阶跃检测:比较掩膜内侧/外侧小窗的 V 均值。
纸边是真实的明暗台阶(阶跃大),木纹/阴影纹理阶跃小。相比固定亮度
阈值(V>0.60 / V<0.45),不受整体曝影响。两级窗口:掩膜准确的
位置用 r_small,掩膜略有内收/偏移处依靠 r_big 补上。
"""
_, V = rgb_to_sv(arr)
p = paper.astype(np.float32)
q = 1.0 - p
def side_means(r):
din = box_mean(p, r)
dout = box_mean(q, r)
vin = box_mean(V * p, r) / np.maximum(din, 1e-6)
vout = box_mean(V * q, r) / np.maximum(dout, 1e-6)
ok = (din > 0.08) & (dout > 0.08)
return np.where(ok, np.abs(vin - vout), 0.0)
step = np.maximum(side_means(r_small), side_means(r_big))
return step > step_th
# ---------------- 四角提取(掩膜 -> 四边拟合 -> 交点) ----------------
def mask_boundary_points(mask):
"""掩膜边界像素坐标 (N,2) float32,顺序为 (x, y)"""
inner = erode(mask, 3)
ys, xs = np.where(mask & ~inner)
return np.stack([xs, ys], axis=1).astype(np.float32)
def _fit_line(pts):
"""总体最小二乘拟直线,返回法向量 n 与 n·p = c"""
c0 = pts.mean(axis=0)
q = pts - c0
cov = (q.T @ q) / max(len(pts) - 1, 1)
w, v = np.linalg.eigh(cov)
d = v[:, int(np.argmax(w))] # 主方向
n = np.array([-d[1], d[0]], dtype=np.float64)
nn = np.linalg.norm(n)
if nn < 1e-9:
return np.array([0.0, 1.0]), float(c0[1])
n = n / nn
return n, float(n @ c0)
def _intersect(n1, c1, n2, c2):
A = np.vstack([n1, n2])
b = np.array([c1, c2], dtype=np.float64)
try:
return np.linalg.solve(A, b)
except np.linalg.LinAlgError:
return None
def detect_quad_corners(paper, points=None, robust_px=2.0, group_tol=22.0):
"""提取四个纸角,返回 [(x,y)],顺序 左上、右上、右下、左下。
锚点一律取自纸面掩膜边界(稳定,不受噪声小刺影响);points 传入纸边黑线
像素时,按到四条基准边的距离分组后重新拟合,得到更贴真实纸边的角点。
每个角都有退化与越界检查,不通过就回退到掩膜基准边交点。
"""
base = mask_boundary_points(paper)
if len(base) < 40:
return []
bx, by = base[:, 0], base[:, 1]
a_s = bx + by
a_d = bx - by
anchors = np.array([
base[int(np.argmin(a_s))], # 左上
base[int(np.argmax(a_d))], # 右上
base[int(np.argmax(a_s))], # 右下
base[int(np.argmin(a_d))], # 左下
], dtype=np.float64)
base_lines = []
for a, b in ((0, 1), (1, 2), (2, 3), (3, 0)):
p0, p1 = anchors[a], anchors[b]
dv = p1 - p0
L = float(np.linalg.norm(dv))
if L < 1e-6:
base_lines.append(None)
continue
dv = dv / L
n0 = np.array([-dv[1], dv[0]])
base_lines.append((n0, float(n0 @ p0)))
corners = []
for i in range(4):
l1, l2 = base_lines[i - 1], base_lines[i]
if l1 is None or l2 is None:
corners.append(anchors[i].copy())
continue
p = _intersect(l1[0], l1[1], l2[0], l2[1])
corners.append(anchors[i].copy() if p is None else p)
if points is None:
return [(float(p[0]), float(p[1])) for p in corners]
pts = np.asarray(points, dtype=np.float64)
if len(pts) < 40:
return [(float(p[0]), float(p[1])) for p in corners]
dd = []
for l in base_lines:
if l is None:
dd.append(np.full(len(pts), 1e9))
else:
dd.append(np.abs(pts @ l[0] - l[1]))
D = np.stack(dd, axis=1)
idx = np.argmin(D, axis=1)
dmin = D[np.arange(len(pts)), idx]
lines = []
for i in range(4):
sel = pts[(idx == i) & (dmin <= group_tol)]
if len(sel) < 40:
lines.append(base_lines[i])
continue
n, c = _fit_line(sel)
for _ in range(2):
res = np.abs(sel @ n - c)
keep = sel[res <= max(robust_px, 1.0) * 2.5]
if len(keep) < 40:
break
n, c = _fit_line(keep)
sel = keep
lines.append((n, c))
H, W = paper.shape
for i in range(4):
l1, l2 = lines[i - 1], lines[i]
if l1 is None or l2 is None:
continue
det = l1[0][0] * l2[0][1] - l2[0][0] * l1[0][1]
if abs(det) < 1e-3: # 两边近共线,交点发散
continue
p = _intersect(l1[0], l1[1], l2[0], l2[1])
if p is None:
continue
if not (-0.15 * W <= p[0] <= 1.15 * W and -0.15 * H <= p[1] <= 1.15 * H):
continue
corners[i] = p
return [(float(p[0]), float(p[1])) for p in corners]
# ---------------- 覆盖率指标 ----------------
def border_coverage_exact(edges, paper, tol=14):
"""沿掩膜逐行/逐列的真实边界测覆盖率(适配斜边,不受 bbox 影响)。
对每一列取掩膜的上/下边界,对每一行取左/右边界,检查 ±tol 内是否有边缘像素。
"""
H, W = paper.shape
col_has = paper.any(axis=0)
row_has = paper.any(axis=1)
top_hit = top_n = 0
bot_hit = bot_n = 0
for x in np.where(col_has)[0]:
ys = np.where(paper[:, x])[0]
y0, y1 = ys.min(), ys.max()
if edges[max(0, y0 - tol):min(H, y0 + tol + 1), x].any():
top_hit += 1
top_n += 1
if edges[max(0, y1 - tol):min(H, y1 + tol + 1), x].any():
bot_hit += 1
bot_n += 1
left_hit = left_n = 0
right_hit = right_n = 0
for y in np.where(row_has)[0]:
xs = np.where(paper[y, :])[0]
x0, x1 = xs.min(), xs.max()
if edges[y, max(0, x0 - tol):min(W, x0 + tol + 1)].any():
left_hit += 1
left_n += 1
if edges[y, max(0, x1 - tol):min(W, x1 + tol + 1)].any():
right_hit += 1
right_n += 1
return {
"top": top_hit / max(top_n, 1),
"bottom": bot_hit / max(bot_n, 1),
"left": left_hit / max(left_n, 1),
"right": right_hit / max(right_n, 1),
}
# ---------------- 对外接口 ----------------
def detect(image_path, max_side=1600):
"""检测纸张四角,返回 (corners_原图坐标, info)。失败返回 ([], info)。"""
src = ImageOps.exif_transpose(Image.open(image_path)).convert('RGB')
src_w, src_h = src.size
pil, arr = load_and_scale(image_path, max_side=max_side)
H, W = arr.shape[:2]
paper = paper_mask(arr)
frac = float(paper.sum()) / (H * W)
band = boundary_band(paper, _odd(max(15, H // 45)))
edges = edge_map(arr)
step = boundary_step_filter(arr, paper)
cand = closing(edges & band & step, _odd(max(5, H // 160)))
cand = opening(cand, _odd(max(3, H // 500)))
ey, ex = np.nonzero(cand)
pts = np.stack([ex, ey], axis=1).astype(np.float32) if len(ex) > 40 else None
corners = detect_quad_corners(paper, points=pts)
scale = min(1.0, float(max_side) / max(src_w, src_h))
src_pts = [(x / scale, y / scale) for x, y in corners] if corners else []
info = {
'source_size': [src_w, src_h],
'processed_size': [W, H],
'scale': round(scale, 6),
'paper_frac': round(frac, 4),
'edge_pixels': int(cand.sum()),
'coverage': border_coverage_exact(cand, paper, tol=max(10, H // 110)) if corners else {},
'processed_pts': corners,
'edge_mask_small': cand, # 处理尺寸上的边线掩膜(供叠加出图)
}
return src_pts, info