From 847ca3cd206657d875abb82a565da632442ca1f4 Mon Sep 17 00:00:00 2001 From: ljq Date: Tue, 27 Oct 2026 14:31:33 +0800 Subject: [PATCH] Configure macOS native OCR and document mydsv41f vision --- memory/ocr_utils.py | 113 +++++++++++++++++++++++++++++++++++-------- memory/vision_sop.md | 14 +++++- 2 files changed, 105 insertions(+), 22 deletions(-) diff --git a/memory/ocr_utils.py b/memory/ocr_utils.py index e296ed48..8f1ddb8e 100644 --- a/memory/ocr_utils.py +++ b/memory/ocr_utils.py @@ -1,13 +1,18 @@ """ 本地 OCR 工具 -- OCR引擎: rapidocr-onnxruntime (~1s/次, 中英文准确率高, 带bbox) +- OCR引擎: macOS默认Apple Vision（pyobjc-framework-Vision）；其他平台rapidocr-onnxruntime +- 输出兼容text/lines/details；bbox为输入图左上原点四角像素，enhance后自动缩回 +- macOS截图走macljqCtrl（需要录屏权限）；图片文件OCR不需要录屏权限 - 坑(rapid): result[i][2] conf 是 str 不是 float - 坑(rapid): 无文字时 result 返回 None 而非空列表 - 坑: enhance 放大+高对比度处理，对清晰文字有害，默认关闭 - 坑(远程桌面): ImageGrab/mss 在 RDP 断开后截图全黑，用 ocr_window(hwnd) 代替 """ import re -from PIL import ImageGrab, Image, ImageEnhance +import sys +import os +from io import BytesIO +from PIL import ImageGrab, Image, ImageEnhance, ImageOps _LANG = 'zh-Hans-CN' _rapid_engine = None @@ -39,38 +44,106 @@ def _ocr_rapid(img): text = _strip_cjk_spaces('\n'.join(lines)) return {'text': text, 'lines': [_strip_cjk_spaces(l) for l in lines], 'details': details} +def _ocr_apple(img, lang): + import objc + import Vision + import Quartz + from Foundation import NSData + aliases = {'zh-Hans-CN': 'zh-Hans', 'zh-CN': 'zh-Hans', 'en': 'en-US'} + languages = list(lang) if isinstance(lang, (list, tuple)) else [lang or 'zh-Hans'] + languages = [aliases.get(l, l) for l in languages] + if 'en-US' not in languages: + languages.append('en-US') + with objc.autorelease_pool(): + buf = BytesIO() + img.save(buf, format='PNG') + raw = buf.getvalue() + data = NSData.dataWithBytes_length_(raw, len(raw)) + source = Quartz.CGImageSourceCreateWithData(data, None) + cg = Quartz.CGImageSourceCreateImageAtIndex(source, 0, None) + request = Vision.VNRecognizeTextRequest.alloc().init() + request.setRecognitionLevel_(Vision.VNRequestTextRecognitionLevelAccurate) + request.setRecognitionLanguages_(languages) + request.setUsesLanguageCorrection_(True) + handler = Vision.VNImageRequestHandler.alloc().initWithCGImage_options_(cg, {}) + ok, error = handler.performRequests_error_([request], None) + if not ok: + raise RuntimeError(f'Apple Vision OCR failed: {error}') + details = [] + w, h = img.size + for observation in request.results() or []: + candidates = observation.topCandidates_(1) + if not candidates: + continue + candidate = candidates[0] + box = observation.boundingBox() + # Apple: normalized bottom-left -> image pixel top-left, rapid-compatible quad. + x1 = max(0, min(w, box.origin.x * w)) + y1 = max(0, min(h, (1 - box.origin.y - box.size.height) * h)) + x2 = max(0, min(w, (box.origin.x + box.size.width) * w)) + y2 = max(0, min(h, (1 - box.origin.y) * h)) + details.append({'bbox': [[x1, y1], [x2, y1], [x2, y2], [x1, y2]], + 'text': _strip_cjk_spaces(str(candidate.string())), + 'conf': float(candidate.confidence())}) + lines = [d['text'] for d in details] + return {'text': '\n'.join(lines), 'lines': lines, 'details': details} + + def ocr_image(image_input, lang=_LANG, enhance=False, engine=None): + """PIL/path -> {text, lines, details}; bbox为输入图左上原点像素四角。 + engine=None: macOS Apple Vision，其他平台rapid；可指定apple/vision/rapid。 + lang支持语言字符串或列表（Apple后端）；enhance默认关闭。 """ - 对 PIL Image 做 OCR - :param image_input: PIL Image 对象 或 文件路径(str) - :param lang: 保留参数，当前未使用 - :param enhance: 预处理 - :param engine: 保留参数，当前仅支持 rapid/None - :return: dict {'text': 全文, 'lines': [行文本], 'details': [bbox+conf]} - """ - if isinstance(image_input, str): - image_input = Image.open(image_input) + if isinstance(image_input, (str, os.PathLike)): + with Image.open(image_input) as source: + image_input = ImageOps.exif_transpose(source).convert('RGB') + elif isinstance(image_input, Image.Image): + image_input = ImageOps.exif_transpose(image_input).convert('RGB') + else: + raise TypeError('image_input must be a path or PIL Image') + original_size = image_input.size if enhance: image_input = _preprocess(image_input) - if engine not in (None, 'rapid'): - raise ValueError("Only rapid OCR is supported") - return _ocr_rapid(image_input) + engine = engine or ('apple' if sys.platform == 'darwin' else 'rapid') + if engine in ('apple', 'vision'): + if sys.platform != 'darwin': + raise RuntimeError('Apple Vision requires macOS') + result = _ocr_apple(image_input, lang) + elif engine == 'rapid': + result = _ocr_rapid(image_input) + else: + raise ValueError('engine must be apple, vision, rapid or None') + if enhance: + sx, sy = original_size[0] / image_input.width, original_size[1] / image_input.height + for detail in result['details']: + detail['bbox'] = [[float(x) * sx, float(y) * sy] for x, y in detail['bbox']] + return result def ocr_screen(bbox=None, lang=_LANG, enhance=False, engine=None): """ 截取屏幕区域并 OCR - :param bbox: (x1, y1, x2, y2) 像素坐标，None=全屏 - :return: dict {'text': 全文, 'lines': [行文本], 'details': [bbox+conf](仅rapid)} + :param bbox: 必填(x1, y1, x2, y2)物理像素坐标，不允许全屏 + :return: dict，bbox相对截图左上角 """ - img = ImageGrab.grab(bbox=bbox) + if bbox is None or len(bbox) != 4 or bbox[2] <= bbox[0] or bbox[3] <= bbox[1]: + raise ValueError('Provide a valid region bbox=(left, top, right, bottom)') + if sys.platform == 'darwin': + import macljqCtrl as ctrl + img = ctrl.GrabScreen(bbox) + else: + # Preserve the existing non-macOS region-capture implementation. + img = ImageGrab.grab(bbox=bbox) return ocr_image(img, lang, enhance, engine) def ocr_window(hwnd, lang=_LANG, enhance=False, engine=None): """ - 截取窗口并 OCR (使用 PrintWindow API，支持远程桌面断开场景) - :param hwnd: 窗口句柄(int) - :return: dict {'text': 全文, 'lines': [行文本], 'details': [bbox+conf](仅rapid)} + 截取窗口并 OCR (macOS: macljqCtrl；Windows: PrintWindow支持远程桌面断开) + :param hwnd: macOS窗口ID或Windows窗口句柄(int) + :return: dict，bbox相对窗口截图左上角 """ + if sys.platform == 'darwin': + import macljqCtrl as ctrl + return ocr_image(ctrl.GrabWindow(hwnd), lang, enhance, engine) import win32gui, win32ui from ctypes import windll l, t, r, b = win32gui.GetWindowRect(hwnd) diff --git a/memory/vision_sop.md b/memory/vision_sop.md index 83a3d906..a4aa065a 100644 --- a/memory/vision_sop.md +++ b/memory/vision_sop.md @@ -2,7 +2,7 @@ ## ⚠️ 前置规则（必须遵守） -1. **先枚举窗口**：调用 vision 前必须先用 `pygetwindow` 枚举窗口标题，确认目标窗口存在且已激活到前台。窗口不存在就不要截图。 +1. **截图前先枚举窗口**：macOS用 `macljqCtrl.ListWindows()` / `ActivateApp()`，Windows用 `pygetwindow`；确认目标窗口存在且已激活到前台。已有图片文件无需枚举窗口。 2. **🚫 禁止全屏截图**：必须先利用ljqCtrl截取窗口区域。能截局部（如标题栏）就不截整窗口，能截窗口就绝不全屏。全屏截图在任何场景下都不允许。 3. **能不用 vision 就不用**：如果窗口标题/本地 OCR（`ocr_utils.py`）能获取所需信息，就不要调用 vision API，省 token 且更可靠。Vision 是最后手段。 @@ -13,11 +13,21 @@ from vision_api import ask_vision result = ask_vision(image, prompt="描述图片内容", timeout=60, max_pixels=1_440_000) # image: 文件路径(str/Path) 或 PIL Image # backend: 'openai'(默认) | 'claude' | 'modelscope' | 'auto' -# 默认配置: native_oai_config0 (xAI，使用配置内代理)；文件路径/PIL Image 均已实测 +# 默认配置: native_oai_config_dsv4f_11 (mydsv41f / deepseek-v41-flash) +# 已实测文字图和无文字彩色几何图；旧 native_oai_config_dsv4f 测试返回“没有图片”，勿混用 +# 旧xAI配置 native_oai_config0 可手动切回（代理随配置） # 旧配置 native_oai_config15 返回 HTTP 503；备用 native_oai_config_fr113 曾实测识图成功 # 返回 str：成功为模型回复，失败为 'Error: ...' ``` +## 本地 OCR（macOS） + +- `ocr_utils.ocr_image(path_or_pil)` 默认使用Apple Vision，本地处理；依赖 `pyobjc-framework-Vision`。 +- 已实测中英文、数字、空白图、Path/PIL输入及enhance坐标回缩；返回 `text/lines/details`，details含 `bbox/text/conf`。 +- bbox是输入图左上原点的四角像素坐标，不能直接当作屏幕绝对坐标。 +- `ocr_screen(bbox)` 必填物理像素区域；`ocr_window(window_id)` 走 `macljqCtrl.GrabWindow`。窗口/区域截图需要宿主录屏授权，文件OCR不需要。 +- 非macOS默认仍为rapid，也可显式 `engine='rapid'`。字体测试需真实中文字体，PIL默认小字体会产生误导结果。 + ## 如果没有 `vision_api.py`，初次构建vision能力 1. 复制 `memory/vision_api.template.py` → `memory/vision_api.py`