158 lines
3.7 KiB
Markdown
158 lines
3.7 KiB
Markdown
|
---
|
|||
|
|
name: openakita/skills@image-understanding
|
|||
|
|
description: Analyze images using Dashscope (Qwen) Vision models. Supports detailed description, OCR text extraction, object recognition, and visual Q&A. When you need to understand image content using Alibaba Cloud Dashscope API.
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# 图片理解技能 (Image Understanding)
|
|||
|
|
|
|||
|
|
使用 **Dashscope（通义千问）** 视觉模型分析图片，支持详细描述、OCR文字提取、物体识别和图片问答。
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 简介
|
|||
|
|
|
|||
|
|
图片理解技能是一个强大的视觉分析工具，通过调用 Dashscope（阿里云通义千问）的视觉大模型（qwen-vl-plus、qwen-vl-max），让 AI 能够理解和分析图像内容。
|
|||
|
|
|
|||
|
|
**核心功能：**
|
|||
|
|
- 🖼️ 图片内容详细描述
|
|||
|
|
- 🔤 文字提取（OCR）
|
|||
|
|
- 🎯 物体识别
|
|||
|
|
- 💬 图片问答
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 使用场景
|
|||
|
|
|
|||
|
|
### 📄 文档处理
|
|||
|
|
- 会议白板照片转文字
|
|||
|
|
- 纸质文档扫描识别
|
|||
|
|
- 手写笔记数字化
|
|||
|
|
|
|||
|
|
### 🛒 工作应用
|
|||
|
|
- 产品图片分析
|
|||
|
|
- 竞品图片提取信息
|
|||
|
|
- 图表数据解读
|
|||
|
|
|
|||
|
|
### 💬 图片问答
|
|||
|
|
- 针对图片提问获取答案
|
|||
|
|
- 理解复杂场景细节
|
|||
|
|
- 技术图纸逻辑分析
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 环境配置
|
|||
|
|
|
|||
|
|
### 1️⃣ 安装依赖
|
|||
|
|
|
|||
|
|
```bash
|
|||
|
|
pip install requests
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### 2️⃣ 获取 Dashscope API Key
|
|||
|
|
|
|||
|
|
1. 访问 [Dashscope 控制台](https://dashscope.console.aliyun.com/)
|
|||
|
|
2. 创建账号并开通服务
|
|||
|
|
3. 创建 API Key
|
|||
|
|
|
|||
|
|
### 3️⃣ 配置 API Key
|
|||
|
|
|
|||
|
|
```bash
|
|||
|
|
# 方式一：环境变量（推荐）
|
|||
|
|
set DASHSCOPE_API_KEY=sk-your-api-key-here
|
|||
|
|
|
|||
|
|
# 方式二：运行时传入（见下方）
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 使用方法
|
|||
|
|
|
|||
|
|
### 基本命令
|
|||
|
|
|
|||
|
|
```bash
|
|||
|
|
python scripts/image_understanding.py -i 图片路径 [选项]
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### 常用参数
|
|||
|
|
|
|||
|
|
| 参数 | 说明 |
|
|||
|
|
|------|------|
|
|||
|
|
| `-i, --image` | **必填** 图片路径或URL |
|
|||
|
|
| `-m, --model` | 模型选择：`qwen-vl-plus`(默认) 或 `qwen-vl-max` |
|
|||
|
|
| `-p, --custom-prompt` | 自定义分析提示词 |
|
|||
|
|
| `-e, --extract-text` | 提取文字(OCR) |
|
|||
|
|
| `-o, --identify-objects` | 识别物体 |
|
|||
|
|
| `--compact` | 输出紧凑JSON |
|
|||
|
|
|
|||
|
|
### 使用示例
|
|||
|
|
|
|||
|
|
```bash
|
|||
|
|
# 1. 基本描述（默认）
|
|||
|
|
python scripts/image_understanding.py -i photo.jpg
|
|||
|
|
|
|||
|
|
# 2. 提取文字
|
|||
|
|
python scripts/image_understanding.py -i screenshot.png -e
|
|||
|
|
|
|||
|
|
# 3. 识别物体
|
|||
|
|
python scripts/image_understanding.py -i photo.jpg -o
|
|||
|
|
|
|||
|
|
# 4. 自定义问答
|
|||
|
|
python scripts/image_understanding.py -i photo.jpg -p "这个产品多少钱？"
|
|||
|
|
|
|||
|
|
# 5. 使用更强的模型
|
|||
|
|
python scripts/image_understanding.py -i photo.jpg -m qwen-vl-max
|
|||
|
|
|
|||
|
|
# 6. 网络图片
|
|||
|
|
python scripts/image_understanding.py -i "https://example.com/image.png" -e
|
|||
|
|
|
|||
|
|
# 7. 设置API Key后运行
|
|||
|
|
set DASHSCOPE_API_KEY=sk-xxx
|
|||
|
|
python scripts/image_understanding.py -i photo.jpg
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 最佳实践
|
|||
|
|
|
|||
|
|
### 📸 图片质量
|
|||
|
|
- 确保图片清晰、亮度充足
|
|||
|
|
- 文字图片分辨率不低于 640x640
|
|||
|
|
- 避免模糊或过暗的图片
|
|||
|
|
|
|||
|
|
### 💡 提示词技巧
|
|||
|
|
- 使用具体、明确的指令
|
|||
|
|
- 指定关注点（如"重点关注价格标签"）
|
|||
|
|
- 多语言场景可混合中英文
|
|||
|
|
|
|||
|
|
### ✅ 结果验证
|
|||
|
|
- 重要信息建议人工复核
|
|||
|
|
- 涉及专业领域需专家确认
|
|||
|
|
- 妥善保存原始图片和分析结果
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## API 配置
|
|||
|
|
|
|||
|
|
| 配置项 | 值 |
|
|||
|
|
|--------|-----|
|
|||
|
|
| 服务商 | Dashscope (通义千问) |
|
|||
|
|
| 默认模型 | qwen-vl-plus |
|
|||
|
|
| 高级模型 | qwen-vl-max |
|
|||
|
|
| API Base | `https://dashscope.aliyuncs.com/compatible-mode/v1` |
|
|||
|
|
| 环境变量 | `DASHSCOPE_API_KEY` |
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 故障排除
|
|||
|
|
|
|||
|
|
| 问题 | 解决方案 |
|
|||
|
|
|------|----------|
|
|||
|
|
| API Key 错误 | 检查 `DASHSCOPE_API_KEY` 是否正确 |
|
|||
|
|
| 图片格式不支持 | 使用 PNG/JPG/GIF/WEBP/BMP 格式 |
|
|||
|
|
| 网络超时 | 检查网络连接，尝试使用代理 |
|
|||
|
|
| 识别不准确 | 提高图片质量，添加更详细的提示词 |
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
运行 `python scripts/image_understanding.py --help` 查看完整帮助
|