code_hw_object_v2 / code /rule_16.py
zry-research's picture
feat: add weight to model/
3bda1a9
Raw History Blame
8.85 kB
import os, json
import numpy as np
from pathlib import Path
from typing import Iterable, Optional, Union
from ultralytics import YOLO
import re
promotion_keywords = [
r"领", r"申请", r"立即",r"马上",r"即刻",r"立即",r"下载",
r"一定",r"现发",
r"大牌美食",
]
extra_keywords = [
r"立即领取",r"点击领取",r"申请我的额度",r"免费观看",
r"火热选购",
r"立即前往",r"立即下载",r"立即投保",r"立即参与",r"立即解救",r"立即领取",r"立即抢购",r"立即购买",r"立即签到",
r"测一测",r"一定要买",r"现摘现发",
r"登录",r"上滑",
]
compiled_patterns = [re.compile(p, flags=re.IGNORECASE) for p in (promotion_keywords + extra_keywords)]
compiled_extra_patterns = [re.compile(p, flags=re.IGNORECASE) for p in extra_keywords]
def extract_texts(ocr_result) -> str:
# 支持 str / [str] / dict / list[dict/...]
texts = []
texts = ocr_result[0]['rec_texts']
scores = ocr_result[0]['rec_scores']
coordss = ocr_result[0]['rec_polys']
ocr_result = [
(pts, (txt, conf))
for pts, txt, conf in zip(coordss, texts, scores)
]
for res in ocr_result:
coords, (text, confidence) = res
if confidence > 0.5:
texts.append(text)
merged = " ".join([t for t in texts if t]).strip()
merged = re.sub(r"\s+", "", merged)
print(merged)
return merged
def merge_by_rows(ocr_result, y_threshold=40):
blocks = []
texts = ocr_result[0]['rec_texts']
scores = ocr_result[0]['rec_scores']
coordss = ocr_result[0]['rec_polys']
ocr_result = [
(pts, (txt, conf))
for pts, txt, conf in zip(coordss, texts, scores)
]
for pts, (txt, conf) in ocr_result:
if conf < 0.5 or not txt.strip():
continue
ys = [p[1] for p in pts]
xs = [p[0] for p in pts]
blocks.append({
"text": txt.strip(),
"x_min": min(xs),
"y_max": max(ys)
})
if not blocks:
return ""
blocks.sort(key=lambda b: b["y_max"])
lines, cur = [], [blocks[0]]
for blk in blocks[1:]:
if abs(blk["y_max"] - cur[0]["y_max"]) <= y_threshold:
cur.append(blk)
else:
lines.append(cur)
cur = [blk]
lines.append(cur)
merged = []
for line in lines:
line.sort(key=lambda b: b["x_min"])
merged.append("".join(b["text"] for b in line))
return " ".join(merged)
# ========== 依赖函数(全部保留,原样引入即可) ==========
def is_guide_button_with_yolo_image(
image_np: np.ndarray,
model: Union[str, YOLO],
ocr,
out_dir: Union[str, Path] = "yolo_outputs",
name: str = "predict",
imgsz: int = 1024,
device: Union[int, str] = 1,
iou: float = 0.45,
max_det: int = 300,
half: bool = False,
verbose: bool = True,
idx: Optional[Union[int, str]] = None # 新增:指定保存文件名
) -> bool:
"""
用 YOLO 对单张 numpy 图像进行检测,并保存可视化与 JSON。
返回:是否包含促销元素(True/False)
"""
class_thresholds = {
0: 1.0, # 优惠券
1: 0.75, # 促销贴纸
2: 1.0, # 元宝
3: 1.0, # 大礼物盒子
4: 0.95, # 爆炸
5: 0.95, # 礼物盒子
6: 0.95, # 福袋
7: 0.95, # 红包
8: 0.95, # 金币
9: 0.95, # 金币簇
10: 0.35, # 长条促销贴纸
11: 0.35, # 长条贴纸
}
# 加载模型
if isinstance(model, str):
model = YOLO(model)
out_dir = Path(out_dir)
out_dir.mkdir(parents=True, exist_ok=True)
results = model.predict(
source=[image_np],
imgsz=imgsz,
conf=0.25,
iou=iou,
device=device,
max_det=max_det,
half=half,
save=True,
project=str(out_dir),
name=name,
exist_ok=True,
verbose=verbose
)
promo_found = False
det_list = []
names_dict = getattr(model, "names", None)
for result in results:
save_dir = Path(result.save_dir)
if result.boxes is None:
continue
boxes = result.boxes.xyxy.cpu().numpy()
classes = result.boxes.cls.cpu().numpy().astype(int)
scores = result.boxes.conf.cpu().numpy()
if verbose:
print("—— detections ——")
h, w = image_np.shape[:2]
total_img_area = h * w # 图像总面积
for (x1, y1, x2, y2), cid, score in zip(boxes, classes, scores):
cname = names_dict.get(int(cid), str(int(cid))) if isinstance(names_dict, dict) else str(int(cid))
if verbose:
print(f"cls={cname}({cid}) conf={score:.3f}")
det_list.append([cname])
det_list.append([float(x1), float(y1), float(x2), float(y2), int(cid), float(score)])
th = class_thresholds.get(cid, None)
if th is None:
continue
# ================= 修改部分开始 =================
# 处理非长条类型的类别 (即 0-9)
if cid not in (10, 11):
if cid == 1:
# 计算当前检测框的面积
box_area = (x2 - x1) * (y2 - y1)
# 计算占比
ratio = box_area / total_img_area if total_img_area > 0 else 0
# 逻辑:如果是类别1,置信度达标 且 面积占比小于 0.01,判定为True
if score >= th and ratio < 0.01:
if verbose:
print(f" [Match] Class 1 (Area Ratio: {ratio:.4f} < 0.01)")
promo_found = True
else:
# 其他类别 (0, 2-9) 保持原逻辑:只看置信度
if score >= th:
promo_found = True
continue
# ================= 修改部分结束 =================
# 下面是处理类别 10, 11 的逻辑 (原代码保持不变)
if cid in (10,11) and score >= 0.875:
promo_found = True
continue
if score >= th:
if ocr is not None:
# 裁剪检测框区域
xi1, yi1 = max(0, int(x1)), max(0, int(y1))
xi2, yi2 = min(w - 1, int(x2)), min(h - 1, int(y2))
if xi2 > xi1 and yi2 > yi1:
crop = image_np[yi1:yi2, xi1:xi2]
try:
ocr_result = ocr.ocr(crop)
# text_merged = extract_texts(ocr_result)
text_merged = merge_by_rows(ocr_result)
print(text_merged)
# 正则匹配:命中任一关键词则判定为促销
if any(p.search(text_merged) for p in compiled_patterns):
if verbose:
print(f"[OCR-MATCH] {cname} 命中促销关键词")
promo_found = True
continue
except Exception as e:
if verbose:
print(f"[OCR-ERROR] {e}")
# OCR 失败时不改变 promo_found,保持谨慎
# —— 重命名保存的图片 ——
if idx is not None:
default_path = save_dir / "image0.jpg"
target_path = save_dir / f"{idx}.jpg"
if default_path.exists():
default_path.rename(target_path)
if verbose:
print(f"[保存] 可视化检测图: {target_path}")
# 保存检测结果 JSON
json_path = save_dir / f"detections_{idx if idx is not None else '0'}.json"
with open(json_path, "w", encoding="utf-8") as f:
json.dump(det_list, f, ensure_ascii=False, indent=2)
if verbose:
print(f"[保存] 检测数组: {json_path}")
# ==== 检测框全部没命中,再做一次全局OCR,匹配extra_keywords ====
if ocr is not None and not promo_found:
try:
global_ocr_result = ocr.ocr(image_np)
global_text = merge_by_rows(global_ocr_result)
print("[全局OCR]", global_text)
if any(p.search(global_text) for p in compiled_extra_patterns):
if verbose:
print("[全局OCR-MATCH] 命中extra_keywords促销关键词")
return True
except Exception as e:
if verbose:
print(f"[全局OCR-ERROR] {e}")
return promo_found