| 功能 | 说明 |
|---|---|
| 🎯 划选即识别 | 点击图标 → 屏幕冻结 → 拖拽划选 → 毫秒级 OCR 识别 |
| 🔍 文字块筛选 | Filter 模式下点击图像上的文字块，选择/排除特定行 |
| 🌊 WebGL 玻璃效果 | 3D 流体扭曲 + 色散 + 霓虹选区 + 暗化背景 |
| 🖥️ 赛博 Data Pad | 军工级终端风格结果面板（LED · HEX · 扫描线 · 呼吸灯动画） |
| 🎨 多主题 & i18n | 5 套识别主题 + 3 套控制台主题，中英双语自动适配 |
| ☁️ 按需下载模型 | 首次识别自动拉取（约 22MB），永久缓存至 IndexedDB |
| 🔒 完全本地推理 | 所有 OCR 在浏览器内完成，数据不离开用户设备 |
| 📦 轻量化包分发 | 纯前端轻量安装（首次使用按需缓存），支持本地模型离线手动导入 |
Content Script (Orchestrator)
├── state.ts ← 集中状态管理 (OcrItem[], activeTab)
├── selection.ts ← 鼠标划选 + 裁剪 + OCR 请求
├── interaction-layer.ts ← 文字块渲染 + 点击交互 + DPR 坐标投影
├── data-pad.ts ← 军工级赛博终端 UI
├── overlay.ts ← Shadow DOM 隔离 + UI 层工厂
├── projector.ts ← 坐标系转换 (Physical ↔ CSS ↔ DPR)
├── i18n.ts ← 国际化模块
└── index.ts ← ~100 行轻量协调器（乐观截图预加载）
Background (Service Worker)
└── worker.ts ← 乐观截图（click 即 capture）+ 消息路由 + 历史存储
Offscreen Document
├── ocr-engine.ts ← Offscreen ↔ Sandbox 消息桥（静默生产日志）
└── sandbox/index.ts ← PaddleOCR.js 引擎（Blob URL 生命周期管理）
UI
└── gl-renderer.ts ← WebGL 背景渲染（VRAM 完整回收：detachShader → deleteProgram → lose_context）
Utils
├── db.ts ← IndexedDB（历史记录 + AI 模型 Blob 存储）
└── models.ts ← 模型下载与校验逻辑
| 层 | 技术 |
|---|---|
| AI 引擎 | @paddleocr/paddleocr-js (ONNX Runtime WASM) |
| 核心逻辑 | TypeScript (strict mode) |
| 视觉渲染 | WebGL + GLSL Shaders |
| UI 隔离 | Shadow DOM |
| 存储 | IndexedDB (Raw implementation) |
| 构建 | Vite + @crxjs/vite-plugin |
| CI/CD | GitHub Actions (Release + Pages + Smoke Test) |
| 代码质量 | oxlint · oxfmt · tsc --noEmit · check-try-catch |
OCRMeow deliberately separates runtime code from model data:
| Target | ORT WASM Runtime | OCR Models |
|---|---|---|
| Chrome Web Store / extension build | Bundle the single required local JSEP runtime | Fetch on demand from CDN/Release fallback, cache in IDB |
| Local unpacked extension | Same as Chrome Web Store for parity | Fetch on demand or import manually |
| Web Demo / GitHub Pages | Use the local Pages runtime | Deployed as same-origin Pages assets, cache in IDB |
The extension package keeps ort-wasm-simd-threaded.jsep.wasm local because ORT
WASM is executable runtime code. Remotely downloading it after installation is
risky for Chrome Web Store Manifest V3 review. Model .tar files are data
assets, so they remain fetch-on-demand and are persisted in IndexedDB. GitHub
Release model assets are valid for extension/background fetches, CI fixtures,
and manual downloads, but GitHub Pages browser fetches cannot rely on them
because Release asset redirects do not consistently expose CORS headers. The CI
pipeline therefore creates the Chrome extension ZIP before copying model files
into dist/models/ for the Pages artifact only. Store packages stay lean; the
web demo gets same-origin model downloads.
首次使用 OCRMeow 时，网页 Demo 和 Chrome 扩展都会明确提示用户：OCR 需要先下载约 22MB 的本地模型。模型下载完成后会缓存到当前浏览器的 IndexedDB，后续识别无需重复下载；如果用户正在执行识别，本次请求会在模型准备好后自动继续。
请放心，OCRMeow 会保护用户隐私：所有识别操作都在浏览器本地完成，不会上传到云端。历史记录会按用户设置保存在本地 IndexedDB，内容包括识别文本和源图片 data URL；用户可以在 Settings 中清空历史，或者清除站点数据来同时删除历史与模型缓存。
GitHub Pages 的首次 OCR 初始化可能比扩展环境慢，因为浏览器需要从静态站点加载
ORT runtime、两个 ONNX 模型并完成 PaddleOCR 初始化。Web Studio 会等待更长的
初始化窗口，并将 sandbox 内的初始化/推理超时分别回传为明确错误，避免只显示笼统的
OCR_ENGINE_TIMEOUT。
# 安装依赖
npm install
# 生产构建（内含 check → lint → type-check → build 全链路）
npm run build
# 将 dist/ 目录以"加载已解压的扩展程序"方式载入 Chrome
# 或直接从 GitHub Releases 下载预构建的 zip 包每次提交必须通过以下关卡：
npm run check-try-catch # 禁止 try-catch 滥用
npm run format # oxfmt 代码格式化
npm run lint # oxlint 逻辑缺陷检测
npm run type-check # tsc --noEmit 类型安全审计
node tests/run_ocr_test.js # Puppeteer 冒烟测试
npm run build # Vite 生产构建- Milestone 23 (2026-06-14): Hardened Studio boot so navigation, upload, paste, and settings controls are bound before any asynchronous model status probing or setup wizard work. This prevents slow/failed Pages asset checks from making the dashboard appear frozen, and the OCR smoke test now explicitly asserts History/Settings navigation interactivity.
- Milestone 22 (2026-06-13): Hardened GitHub Pages Web OCR by deploying model
.tarfiles only into the Pages artifact after the lean extension ZIP is already packaged, avoiding browser CORS failures while preserving Chrome Web Store package size. Extended the Web OCR initialization window and added explicit sandbox init/predict timeout errors so slow first-run PaddleOCR startup no longer collapses into a vagueOCR_ENGINE_TIMEOUT. - Milestone 21 (2026-06-10): Pruned the Lean package runtime payload from ~165MB to ~46MB by hiding
public/wasmduring production builds, copying back only the ORT JSEP runtime actually used by PaddleOCR (ort-wasm-simd-threaded.jsep.wasm/.mjs), and deleting Vite's duplicate hashed ORT wasm fallback. Kept ORT wasm local for Chrome Web Store MV3 compliance because it is executable runtime code; model.tarfiles remain fetch-on-demand data. Hardened CI to download real model fixtures from GitHub Releases when local developer model files are absent. History deletion now waits for IndexedDB completion before updating UI. - Milestone 20 (2026-06-10): Canonicalized all public GitHub-facing project links to
OCRMeow. VerifiedOCRMeow/releasesreturns directly while the legacy misspelled releases path redirects, then updated visible model backup links in both Studio entry points to avoid redirect-dependent URLs and future agent confusion. - Milestone 18 (2026-05-30): Eradicated high-priority architectural defects. Replaced stale coordinate caching in
projector.tswith live browser pixel ratio getters to support dynamic page zoom/resolution screen swaps. Prevented global keyboard listener memory leaks in the dashboard lightbox overlay through unified event closures. Declared missing Shadow DOM@keyframesto animate the cyber scanline and led pulses, and introduced an off-screen clipboard fallback for insecure HTTP environments. - Milestone 17 (2026-05-30): Transitioned CI pipeline and build pipeline exclusively to a Lean-Only architecture, avoiding weight bloat by hosting model weight assets in GitHub Release drafts and integrating smart, auto-failover, multi-CDN model syncing gateways.
- Milestone 16 (2026-05-29): Fixed sub-path deployment 404 errors (like GitHub Pages) by switching to relative model fetching paths (
models/det.tar, etc.). Fully unlocked and enabled History Limit controls, Export History, and Clear History on the Web Dashboard using local IndexedDB + JSZip, providing complete feature parity for static web deployments. - Milestone 15 (2026-05-28): Decoupled automatic history saving from the background worker to ensure statelessness. Extracted all CSS from HTML files into a shared Vite asset pipeline, and decomposed the 1380+ line god-file into 7 highly-focused, single-responsibility modules.