古文字 OCR 终于有了统一考场：Chronicles-OCR 开源评测基准解析

# 从 Hugging Face 下载（示例，请替换为实际 repo id）
pip install huggingface_hub
huggingface-cli download Tencent/Chronicles-OCR \
  --repo-type dataset \
  --local-dir ./chronicles_ocr_data \
  --local-dir-use-symlinks False

# 查看数据目录结构
tree ./chronicles_ocr_data -L 2

chronicles_ocr_data/
├── oracle/        # 甲骨文
├── bronze/        # 金文
├── seal/          # 篆书
├── clerical/      # 隶书
├── cursive/       # 草书
├── semi_cursive/  # 行书
├── regular/       # 楷书
└── annotations/   # 标注文件（JSON/CSV）

import json
from pathlib import Path
import matplotlib.pyplot as plt
from PIL import Image

DATA_ROOT = Path("./chronicles_ocr_data")

def load_annotations(script_name: str):
    """加载某一书体的标注文件"""
    ann_path = DATA_ROOT / "annotations" / f"{script_name}.json"
    with open(ann_path, "r", encoding="utf-8") as f:
        return json.load(f)

def show_sample(script_name: str, idx: int = 0):
    """可视化单条样本：原图 + 标注字形 + 现代对应字"""
    ann = load_annotations(script_name)
    sample = ann[idx]
    img_path = DATA_ROOT / script_name / sample["image_file"]
    img = Image.open(img_path)

    fig, ax = plt.subplots(1, 1, figsize=(4, 4))
    ax.imshow(img)
    ax.set_title(
        f"书体: {script_name} | 释读: {sample['modern_char']} | "
        f"原始编号: {sample['char_id']}",
        fontsize=10
    )
    ax.axis("off")
    plt.tight_layout()
    plt.savefig(f"sample_{script_name}_{idx}.png", dpi=150)
    plt.show()

# 示例：查看甲骨文第一条样本
show_sample("oracle", idx=0)

pip install paddleocr paddlepaddle

from paddleocr import PaddleOCR
from pathlib import Path
import json
from collections import defaultdict

# 初始化 OCR 引擎（使用中文模型）
ocr = PaddleOCR(use_angle_cls=True, lang="ch")

DATA_ROOT = Path("./chronicles_ocr_data")

def evaluate_script(script_name: str):
    """对某一书体跑 OCR，逐条比对标注"""
    ann = load_annotations(script_name)
    correct = 0
    total = len(ann)

    for sample in ann:
        img_path = str(DATA_ROOT / script_name / sample["image_file"])
        result = ocr.ocr(img_path, cls=True)

        # 提取 OCR 识别文本
        recognized_chars = []
        for line in result[0]:
            recognized_chars.append(line[1][0])

        # 简化比对：OCR 结果中是否包含目标现代字
        target = sample["modern_char"]
        if target in recognized_chars:
            correct += 1

    acc = correct / total if total > 0 else 0
    return {"script": script_name, "total": total, "correct": correct, "accuracy": acc}

# 七体全跑
scripts = ["oracle", "bronze", "seal", "clerical",
           "cursive", "semi_cursive", "regular"]

results = []
for s in scripts:
    r = evaluate_script(s)
    results.append(r)
    print(f"{r['script']}: {r['accuracy']:.2%} ({r['correct']}/{r['total']})")

# 输出对比表
print("\n=== 七体识别准确率对比 ===")
for r in results:
    print(f"  {r['script']:15s}  {r['accuracy']:.2%}")

# 分析错误模式：哪些字跨书体一致性地识别失败
from collections import Counter

error_chars_by_script = defaultdict(Counter)

for script_name in scripts:
    ann = load_annotations(script_name)
    for sample in ann:
        img_path = str(DATA_ROOT / script_name / sample["image_file"])
        result = ocr.ocr(img_path, cls=True)
        recognized = "".join([line[1][0] for line in result[0]])
        if sample["modern_char"] not in recognized:
            error_chars_by_script[script_name][sample["modern_char"]] += 1

# 打印每个书体 top-5 误识字
for script_name, counter in error_chars_by_script.items():
    print(f"\n{script_name} 最难识别的 5 个字:")
    for char, count in counter.most_common(5):
        print(f"  {char}: 错误 {count} 次")