图片智能处理平台 v2.0
能力矩阵
| 模块 | 功能 | 状态 |
|---|---|---|
| 解析层 | OCR 文字提取 | ✅ |
| 编号智能识别(0/O 混淆修正) | ✅ | |
| 布局结构分析(标题区/内容区/配色区) | ✅ | |
| 感知哈希相似度分组 | ✅ | |
| 理解层 | 图片质量评分(清晰度/对比度/可读性) | ✅ |
| 设计风格分类(科技风/极简风/手绘风) | ✅ | |
| 配色方案提取(主色/辅色/强调色) | ✅ | |
| 元素完整性检测 | ✅ | |
| 生成层 | 豆包生图 API 批量调用 | ✅ |
| Prompt 模板引擎 | ✅ | |
| 生成参数调优 | ✅ | |
| 管控层 | 重复图片自动去重 | ✅ |
| 不合格图片自动打回 | ✅ | |
| 批量重命名标准化 | ✅ | |
| 输出报告生成 | ✅ |
模块 1:图片智能解析引擎
1.1 全区域 OCR 解析
不仅仅读顶部,而是按布局分区读取:
def analyze_layout(img_path):
"""分析图片布局,返回各区域坐标"""
img = Image.open(img_path)
w, h = img.size
return {
"header": (0, 0, w, int(h * 0.25)), # 顶部:编号+标题
"content": (0, int(h*0.25), w, int(h*0.7)), # 中部:内容
"footer": (0, int(h*0.7), w, h), # 底部:色值+领域
}
def ocr_region(img, region_bbox):
"""OCR 读取指定区域"""
crop = img.crop(region_bbox)
crop.save("/tmp/_ocr_region.png")
return subprocess.check_output(["/tmp/ocr", "/tmp/_ocr_region.png"], text=True)
1.2 编号智能识别(解决 0/O 混淆)
def robust_number_detection(ocr_text, img=None):
"""
增强版编号识别,结合字形特征
0 vs O 判定规则:
- 上下文是数字 → 0
- 上下文是英文单词 → O
- 结合图片像素:正圆形 vs 椭圆
"""
import re
# 第一阶段:正则粗提取
candidates = re.findall(r'[Tt]\s*([0-9A-Za-z]{3})', ocr_text)
if not candidates:
return None
raw = candidates[0].upper()
# 第二阶段:字符级修正
correction_map = {
'O': '0', 'o': '0',
'I': '1', 'l': '1',
'Z': '2',
'S': '5',
'B': '8'
}
# 中文 OCR 常见错误(标题修正)
title_correction_map = {
'Al': 'AI',
'丁具': '工具',
'1则化': '正则化',
'多智能休': '多智能体',
'湜示词': '提示词',
}
```
完整的术语知识库和 OCR 错误修正表参见 `ai-flashcard-batch-rename` skill 的 `references/terminology_knowledge_base.md`。
result = []
for c in raw:
result.append(correction_map.get(c, c))
final = ''.join(result)
# 第三阶段:验证必须是纯 3 位数字
if re.match(r'^\d{3}$', final):
return final
return None
1.3 感知哈希相似度分组
解决编号识别失败的图片自动归类:
import imagehash
from collections import defaultdict
def group_by_similarity(image_paths, threshold=0.92):
"""
按图片相似度分组,相同布局/设计的归为一组
threshold: 0.0-1.0,越高越严格
- 0.95: 极严格,几乎完全相同才分组
- 0.92: 默认,实测最佳,布局一致即可
- 0.85: 宽松,风格类似就分组
"""
hashes = {}
for path in image_paths:
h = imagehash.dhash(Image.open(path), hash_size=16)
hashes[path] = h
groups = defaultdict(list)
grouped = set()
for i, path1 in enumerate(image_paths):
if path1 in grouped:
continue
h1 = hashes[path1]
current_group = [path1]
grouped.add(path1)
for path2 in image_paths[i+1:]:
if path2 in grouped:
continue
h2 = hashes[path2]
similarity = 1 - (h1 - h2) / len(h1.hash) ** 2
if similarity >= threshold:
current_group.append(path2)
grouped.add(path2)
groups[f"group_{len(groups):03d}"] = current_group
return dict(groups)
模块 1.4:知识库匹配(标题反推编号)
当 OCR 无法识别编号但能识别标题时,这是提升识别率最有效的步骤(实测 +50%)。
# 标题 -> 编号 映射表(随项目积累扩展)
title_to_num = {
"机器学习": "002",
"深度学习": "003",
"大语言模型": "004",
# ... 70+ 条目(参见 ai-flashcard-batch-rename skill)
}
def match_by_knowledge_base(results, knowledge_base):
"""用知识库补全识别失败的条目"""
for r in results:
if r['final_num'] is None:
title = r['final_title']
# 精确匹配
if title in knowledge_base:
r['final_num'] = knowledge_base[title]
r['num_source'] = 'kb_exact'
# 模糊匹配:前 4 个字一致即可
else:
for known_title, known_num in knowledge_base.items():
if known_title[:4] in title:
r['final_num'] = known_num
r['num_source'] = 'kb_fuzzy'
break
return results
知识库构建策略:
- 先处理 OCR 成功的图片
- 自动提取 标题-编号 对加入知识库
- 每批处理后,知识库自动增长
2.1 质量评分算法
def calculate_quality_score(img_path):
"""
综合质量评分 0-100
维度:清晰度、对比度、文字可读性、色彩一致性
"""
import numpy as np
img = Image.open(img_path)
img_gray = img.convert('L')
arr = np.array(img_gray)
score = 0
# 1. 清晰度(拉普拉斯方差)
from scipy import ndimage
laplacian = ndimage.laplace(arr)
clarity = np.var(laplacian)
clarity_score = min(100, clarity / 50) * 30 # 权重 30%
# 2. 对比度(标准差)
contrast = np.std(arr)
contrast_score = min(100, contrast / 30) * 25 # 权重 25%
# 3. 亮度分布(避免过暗/过亮)
mean_bright = np.mean(arr)
if 80 <= mean_bright <= 200:
brightness_score = 20 # 权重 20%
else:
brightness_score = max(0, 20 - abs(mean_bright - 140) / 3)
# 4. 色彩完整性(检测是否有缺失色块)
color_score = check_color_integrity(img) * 25 # 权重 25%
total_score = clarity_score + contrast_score + brightness_score + color_score
return {
"total": round(total_score, 1),
"clarity": round(clarity_score, 1),
"contrast": round(contrast_score, 1),
"brightness": round(brightness_score, 1),
"color_integrity": round(color_score, 1),
"pass": total_score >= 70
}
def check_color_integrity(img):
"""检测底部色值区域是否完整"""
w, h = img.size
footer = img.crop((0, int(h*0.85), w, h))
# 转换为 HSV 检测纯黑/纯白异常
footer_arr = np.array(footer.convert('HSV'))
sat_mean = np.mean(footer_arr[:, :, 1])
# 饱和度太低说明可能渲染失败
if sat_mean < 20:
return 0.3 # 严重问题
elif sat_mean < 50:
return 0.7 # 轻微问题
return 1.0
2.2 自动质量筛选
def batch_quality_check(directory, threshold=70):
"""批量检测图片质量,分离合格/不合格"""
import shutil
from pathlib import Path
passed_dir = Path(directory) / "_quality_passed"
failed_dir = Path(directory) / "_quality_failed"
passed_dir.mkdir(exist_ok=True)
failed_dir.mkdir(exist_ok=True)
results = []
images = [f for f in os.listdir(directory) if f.endswith('.png')]
for img_file in images:
path = os.path.join(directory, img_file)
score = calculate_quality_score(path)
if score['pass']:
shutil.copy2(path, passed_dir / img_file)
else:
shutil.copy2(path, failed_dir / img_file)
results.append({
"file": img_file,
"score": score
})
# 生成报告
passed = sum(1 for r in results if r['score']['pass'])
report = f"""
📊 图片质量检测报告
━━━━━━━━━━━━━━━━━━━━━━
总图片数: {len(results)}
合格: {passed} ({passed/len(results)*100:.1f}%)
不合格: {len(results)-passed}
平均分: {sum(r['score']['total'] for r in results)/len(results):.1f}
⚠️ 不合格列表:
"""
for r in sorted(results, key=lambda x: x['score']['total'])[:10]:
if not r['score']['pass']:
report += f" {r['file']}: {r['score']['total']} 分\n"
with open(Path(directory) / "quality_report.txt", "w") as f:
f.write(report)
return results
模块 3:配色方案提取
def extract_color_palette(img_path, n_colors=6):
"""提取图片主色调配色方案"""
from sklearn.cluster import KMeans
import numpy as np
img = Image.open(img_path)
img_small = img.resize((100, 100))
arr = np.array(img_small)
# 只取有色彩的像素(排除白色背景)
pixels = arr.reshape(-1, 3)
mask = (pixels.sum(axis=1) < 700) # 排除接近白色
pixels = pixels[mask]
if len(pixels) < 100:
return []
# K-means 聚类
kmeans = KMeans(n_clusters=n_colors, n_init=10, random_state=42)
kmeans.fit(pixels)
colors = kmeans.cluster_centers_.astype(int)
labels = kmeans.labels_
# 按出现频率排序
counts = np.bincount(labels)
sorted_indices = np.argsort(counts)[::-1]
palette = []
for idx in sorted_indices:
r, g, b = colors[idx]
hex_color = f"#{r:02X}{g:02X}{b:02X}"
percentage = counts[idx] / len(labels) * 100
palette.append({
"hex": hex_color,
"rgb": (int(r), int(g), int(b)),
"percentage": round(percentage, 1)
})
return palette
模块 4:批量重命名(升级版)
结合前面所有模块的 smart_rename.py:
# 工作流程
1. 全量 OCR 解析编号和标题
2. 感知哈希分组,解决编号识别失败
3. 质量检测,不合格标记
4. 配色方案提取,附加到文件名元数据
5. 标准化重命名 txxx-标题-版本-质量分.png
6. 生成完整报告
模块 5:图片生成集成
5.1 豆包生图 API 封装
import requests
import json
import base64
from pathlib import Path
class DoubaoImageGenerator:
"""豆包生图 API 封装"""
def __init__(self, api_key, base_url="https://ark.cn-beijing.volces.com/api/v3"):
self.api_key = api_key
self.base_url = base_url
def generate(self, prompt, model="doubao-seed-2-0:260215",
size="1024x1024", n=1, **kwargs):
"""生成图片"""
headers = {
"Authorization": f"Bearer {self.api_key}",
"Content-Type": "application/json"
}
payload = {
"model": model,
"prompt": prompt,
"size": size,
"n": n,
**kwargs
}
response = requests.post(
f"{self.base_url}/images/generations",
headers=headers,
json=payload,
timeout=120
)
if response.status_code == 200:
return response.json()
else:
raise Exception(f"API Error: {response.status_code} - {response.text}")
def batch_generate(self, prompts, output_dir, **kwargs):
"""批量生成"""
output_path = Path(output_dir)
output_path.mkdir(parents=True, exist_ok=True)
results = []
for i, prompt in enumerate(prompts):
try:
result = self.generate(prompt, **kwargs)
# 保存图片
for j, img_data in enumerate(result['data']):
if 'url' in img_data:
# 下载 URL
img_response = requests.get(img_data['url'])
img_path = output_path / f"gen_{i:03d}_{j:02d}.png"
with open(img_path, 'wb') as f:
f.write(img_response.content)
elif 'b64_json' in img_data:
# base64 解码
img_bytes = base64.b64decode(img_data['b64_json'])
img_path = output_path / f"gen_{i:03d}_{j:02d}.png"
with open(img_path, 'wb') as f:
f.write(img_bytes)
results.append({
"prompt_index": i,
"variant": j,
"path": str(img_path),
"success": True
})
except Exception as e:
results.append({
"prompt_index": i,
"error": str(e),
"success": False
})
return results
完整工作流
输入图片目录
↓
[模块 1] 布局解析 + OCR
↓
[模块 1] 编号识别 + 相似度分组补全
↓
[模块 2] 质量检测 + 打分
↓
[模块 3] 配色方案提取
↓
[模块 4] 标准化重命名 txxx-标题-vxx-Qxx.png
↓
[输出] 最终图片 + JSON 报告 + 统计摘要
性能指标
| 操作 | 单张耗时 | 200 张耗时 |
|---|---|---|
| OCR 解析 | 3-5s | 10-15 min |
| 质量检测 | 0.1s | 20s |
| 相似度分组 | 0.05s | 10s |
| 总计 | 3-5s | 10-16 min |
快速开始
# 安装依赖
pip install pillow imagehash numpy scipy scikit-learn requests
# 批量重命名(带质量检测)
python3 scripts/smart_rename.py /path/to/images
# 仅质量检测
python3 scripts/quality_check.py /path/to/images
# 批量生成
python3 scripts/batch_generate.py prompts.txt /output/path
收纳会话档案(2026-10-23 curator 从 batch-file-renaming 移入)
本 skill 的模块 1-4(OCR 直读 → 感知哈希分组 → 知识库匹配 → 批量重命名)与 batch-file-renaming skill 的 AI 名词卡片 3 阶段管线同源。两条实战会话档案已归位于本 skill references/:
references/***SECRET***.md— 华哥 2026-07-13 铁律:必须解图片内容提取编号再重命名,严禁套用 TERM_MAP 预设表(用户否决 lookup-table fallback 的原始会话记录,任何图片识别类任务先读这个)。references/flashcard-session-2026-07-13.md— 79 张 AI 名词卡片重命名会话全记录(manifest 发现、T001-T020 编号规律、4 变体/词条)。
通用管线 SOP 仍在 batch-file-renaming skill(references/ai-flashcard-3stage-pipeline.md + scripts/rename_flashcards.py);本 skill 提供模块级实现与质量检测。