Files
agent_jrxml/backend/annotation_detector.py
T
panda 9bb011e429 feat: v4 multimodal chat input, multi-format support, and annotation detection
- Replace st.chat_input with st-multimodal-chatinput (Ctrl+V paste, drag-drop, file button)
- Extract _process_uploaded_file() shared handler (eliminates ~70 duplicated lines)
- Add XLSX (openpyxl), XLS (xlrd), DOC (olefile) parsers to file_parser.py
- Add backend/annotation_detector.py: circle detection (HoughCircles) + arrow detection (HoughLinesP clustering) + OCR correlation + LLM context formatting
- Add annotation_result field to AgentState with session persistence
- Wire annotation detection into process_input and _format_ocr_context
- Add 11 new tests: 7 annotation detector + 4 multi-format parser
- Update all docs: CLAUDE.md, README.md, CODE_GUIDE.md, ROADMAP.md
2026-05-20 23:43:16 +08:00

332 lines
10 KiB
Python

"""批注检测器:识别图片上的圈选(圆)和箭头,定位用户要修改的字段。
依赖 OpenCV (cv2),从 PaddleOCR 传递依赖已安装。
"""
from __future__ import annotations
import math
from dataclasses import dataclass, field
from typing import Optional
import cv2
import numpy as np
@dataclass
class Annotation:
"""单个批注标记。"""
type: str # "circle" | "arrow"
bbox: dict # {"x": int, "y": int, "w": int, "h": int}
center: tuple[int, int] # (cx, cy)
nearby_texts: list[str] = field(default_factory=list)
from_text: str = "" # 箭头出发点的文本
to_text: str = "" # 箭头指向的文本
from_pt: Optional[tuple[int, int]] = None
to_pt: Optional[tuple[int, int]] = None
def detect_annotations(image_path: str, ocr_elements: list[dict]) -> dict:
"""检测图片上的手写批注(圈选 + 箭头),并与 OCR 文本关联。
Args:
image_path: 图片文件路径
ocr_elements: OCR 元素列表 [{"text": str, "bbox": {x,y,w,h}, "confidence": float}]
Returns:
{"circles": [...], "arrows": [...], "total": int}
"""
img = cv2.imread(image_path)
if img is None:
return {"circles": [], "arrows": [], "total": 0, "error": "无法读取图片"}
h, w = img.shape[:2]
circles = _detect_circles(img)
arrows = _detect_arrows(img)
all_annotations = circles + arrows
_correlate_with_ocr(all_annotations, ocr_elements, w, h)
result: dict = {
"circles": [_annotation_to_dict(a) for a in circles],
"arrows": [_annotation_to_dict(a) for a in arrows],
"total": len(all_annotations),
}
return result
def _annotation_to_dict(a: Annotation) -> dict:
d = {
"type": a.type,
"bbox": a.bbox,
"center": list(a.center),
"nearby_texts": a.nearby_texts,
}
if a.type == "arrow":
d["from_text"] = a.from_text
d["to_text"] = a.to_text
if a.from_pt:
d["from_pt"] = list(a.from_pt)
if a.to_pt:
d["to_pt"] = list(a.to_pt)
return d
# ---------------------------------------------------------------------------
# 圆圈检测
# ---------------------------------------------------------------------------
def _detect_circles(img: np.ndarray) -> list[Annotation]:
"""检测图片中可能是手绘批注的圆圈。"""
h, w = img.shape[:2]
b, g, r = cv2.split(img)
red_enhanced = cv2.addWeighted(r.astype(np.float32), 1.5,
g.astype(np.float32), -0.3, 0)
red_enhanced = cv2.addWeighted(red_enhanced, 1.2,
b.astype(np.float32), -0.3, 0)
red_enhanced = np.clip(red_enhanced, 0, 255).astype(np.uint8)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
combined = cv2.addWeighted(gray, 0.5, red_enhanced, 0.5, 0)
blurred = cv2.GaussianBlur(combined, (9, 9), 2)
min_radius = max(15, min(w, h) // 40)
max_radius = min(200, max(w, h) // 8)
circles_raw = cv2.HoughCircles(
blurred, cv2.HOUGH_GRADIENT, dp=1.2, minDist=min_radius * 2,
param1=50, param2=30, minRadius=min_radius, maxRadius=max_radius,
)
annotations: list[Annotation] = []
if circles_raw is not None:
for cx, cy, r in circles_raw[0]:
bbox = {
"x": max(0, int(cx - r)),
"y": max(0, int(cy - r)),
"w": int(r * 2),
"h": int(r * 2),
}
annotations.append(Annotation(
type="circle",
bbox=bbox,
center=(int(cx), int(cy)),
))
return annotations
# ---------------------------------------------------------------------------
# 箭头检测
# ---------------------------------------------------------------------------
def _detect_arrows(img: np.ndarray) -> list[Annotation]:
"""检测图片中的手绘箭头(直线段 + 端点三角形)。"""
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
edges = cv2.Canny(gray, 50, 150, apertureSize=3)
lines = cv2.HoughLinesP(
edges, rho=1, theta=np.pi / 180, threshold=40,
minLineLength=30, maxLineGap=15,
)
if lines is None:
return []
segments = [(x1, y1, x2, y2) for x1, y1, x2, y2 in lines[:, 0]]
clusters = _cluster_segments(segments)
annotations: list[Annotation] = []
for segs in clusters:
if len(segs) < 2:
continue
all_pts = []
for x1, y1, x2, y2 in segs:
all_pts.append((x1, y1))
all_pts.append((x2, y2))
all_pts_arr = np.array(all_pts)
max_dist = 0
p1 = p2 = all_pts[0]
for i in range(len(all_pts)):
for j in range(i + 1, len(all_pts)):
d = (all_pts[i][0] - all_pts[j][0]) ** 2 + (all_pts[i][1] - all_pts[j][1]) ** 2
if d > max_dist:
max_dist = d
p1, p2 = all_pts[i], all_pts[j]
from_pt, to_pt = _find_arrow_direction(edges, p1, p2)
x1, y1 = from_pt
x2, y2 = to_pt
bbox = {
"x": min(x1, x2),
"y": min(y1, y2),
"w": abs(x2 - x1),
"h": abs(y2 - y1),
}
cx = (x1 + x2) // 2
cy = (y1 + y2) // 2
annotations.append(Annotation(
type="arrow",
bbox=bbox,
center=(cx, cy),
from_pt=from_pt,
to_pt=to_pt,
))
return annotations
def _cluster_segments(segments: list[tuple]) -> list[list[tuple]]:
"""将线段按方向和空间距离聚类。"""
clusters: list[list[tuple]] = []
used = [False] * len(segments)
for i, (x1, y1, x2, y2) in enumerate(segments):
if used[i]:
continue
cluster = [(x1, y1, x2, y2)]
used[i] = True
angle_i = math.atan2(y2 - y1, x2 - x1)
for j in range(i + 1, len(segments)):
if used[j]:
continue
x3, y3, x4, y4 = segments[j]
angle_j = math.atan2(y4 - y3, x4 - x3)
angle_diff = abs(angle_i - angle_j)
if angle_diff > math.pi:
angle_diff = 2 * math.pi - angle_diff
if angle_diff < 0.35:
d1 = math.hypot(x3 - x2, y3 - y2)
d2 = math.hypot(x1 - x4, y1 - y4)
d3 = math.hypot(x3 - x1, y3 - y1)
d4 = math.hypot(x4 - x2, y4 - y2)
if min(d1, d2, d3, d4) < 80:
cluster.append((x3, y3, x4, y4))
used[j] = True
clusters.append(cluster)
return clusters
def _find_arrow_direction(edges: np.ndarray, p1: tuple, p2: tuple) -> tuple[tuple, tuple]:
"""判断箭头的方向(哪端是箭头/三角形汇聚点)。"""
r = 20
h, w = edges.shape[:2]
def edge_density(cx, cy):
x1 = max(0, int(cx - r))
y1 = max(0, int(cy - r))
x2 = min(w, int(cx + r))
y2 = min(h, int(cy + r))
roi = edges[y1:y2, x1:x2]
if roi.size == 0:
return 0
return float(np.count_nonzero(roi)) / roi.size
d1 = edge_density(p1[0], p1[1])
d2 = edge_density(p2[0], p2[1])
if d1 > d2 * 1.3:
return p2, p1
if d2 > d1 * 1.3:
return p1, p2
return p1, p2
# ---------------------------------------------------------------------------
# OCR 关联
# ---------------------------------------------------------------------------
def _correlate_with_ocr(
annotations: list[Annotation],
ocr_elements: list[dict],
img_w: int,
img_h: int,
) -> None:
"""将批注与附近的 OCR 文本关联。"""
if not ocr_elements:
return
for ann in annotations:
ax = ann.center[0]
ay = ann.center[1]
near_texts: list[tuple[str, float]] = []
for elem in ocr_elements:
bbox = elem.get("bbox", {})
ex = bbox.get("x", 0) + bbox.get("w", 0) / 2
ey = bbox.get("y", 0) + bbox.get("h", 0) / 2
dist = math.hypot(ax - ex, ay - ey)
max_dist = max(img_w, img_h) * 0.15
if dist < max_dist:
near_texts.append((elem.get("text", ""), dist))
near_texts.sort(key=lambda x: x[1])
ann.nearby_texts = [t for t, _ in near_texts[:5]]
if ann.type == "arrow" and ann.from_pt and ann.to_pt:
ann.from_text = _closest_text(ann.from_pt, ocr_elements, img_w, img_h)
ann.to_text = _closest_text(ann.to_pt, ocr_elements, img_w, img_h)
def _closest_text(pt: tuple[int, int], ocr_elements: list[dict], img_w: int, img_h: int) -> str:
"""找到离 pt 最近的 OCR 文本。"""
best_text = ""
best_dist = max(img_w, img_h) * 0.12
for elem in ocr_elements:
bbox = elem.get("bbox", {})
ex = bbox.get("x", 0) + bbox.get("w", 0) / 2
ey = bbox.get("y", 0) + bbox.get("h", 0) / 2
dist = math.hypot(pt[0] - ex, pt[1] - ey)
if dist < best_dist:
best_dist = dist
best_text = elem.get("text", "")
return best_text
# ---------------------------------------------------------------------------
# LLM 上下文格式化
# ---------------------------------------------------------------------------
def format_annotation_context(annotation_result: dict) -> str:
"""将批注检测结果格式化为中文 LLM 提示文本。"""
if not annotation_result or not isinstance(annotation_result, dict):
return ""
circles = annotation_result.get("circles", [])
arrows = annotation_result.get("arrows", [])
total = annotation_result.get("total", len(circles) + len(arrows))
if total == 0:
return ""
parts = ["[图片批注检测结果]"]
if circles:
parts.append(f"\n检测到 {len(circles)} 个圈选标记:")
for i, c in enumerate(circles):
center = c.get("center", [0, 0])
near = c.get("nearby_texts", [])
parts.append(
f"{i+1}. 位置 ({center[0]},{center[1]})"
f" — 圈选内容: {', '.join(near) if near else '(附近无文字)'}"
)
if arrows:
parts.append(f"\n检测到 {len(arrows)} 个箭头标记:")
for i, a in enumerate(arrows):
ft = a.get("from_text", "")
tt = a.get("to_text", "")
parts.append(f" 箭头{i+1}. 从「{ft}」→ 指向「{tt}")
parts.append("\n请根据上述圈选/箭头定位用户要修改的报表字段。")
return "\n".join(parts)