""" 本地 OCR 工具 - 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 import sys import os from io import BytesIO from PIL import ImageGrab, Image, ImageEnhance, ImageOps _LANG = 'zh-Hans-CN' _rapid_engine = None def _get_rapid(): global _rapid_engine if _rapid_engine is None: from rapidocr_onnxruntime import RapidOCR _rapid_engine = RapidOCR() return _rapid_engine def _preprocess(img, scale=3, contrast=3.0): img = ImageEnhance.Contrast(img).enhance(contrast) img = img.resize((img.width * scale, img.height * scale)) return img def _strip_cjk_spaces(t): return re.sub(r'(?<=[\u4e00-\u9fff])\s+(?=[\u4e00-\u9fff])', '', t) def _ocr_rapid(img): import numpy as np engine = _get_rapid() arr = np.array(img) result, elapse = engine(arr) if not result: return {'text': '', 'lines': [], 'details': []} lines = [r[1] for r in result] details = [{'bbox': r[0], 'text': r[1], 'conf': float(r[2])} for r in result] 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。 """ 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) 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: 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 """ 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 (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) w, h = r - l, b - t hwndDC = win32gui.GetWindowDC(hwnd) mfcDC = win32ui.CreateDCFromHandle(hwndDC) saveDC = mfcDC.CreateCompatibleDC() saveBitMap = win32ui.CreateBitmap() saveBitMap.CreateCompatibleBitmap(mfcDC, w, h) saveDC.SelectObject(saveBitMap) windll.user32.PrintWindow(hwnd, saveDC.GetSafeHdc(), 3) bmpinfo = saveBitMap.GetInfo() bmpstr = saveBitMap.GetBitmapBits(True) img = Image.frombuffer('RGB', (bmpinfo['bmWidth'], bmpinfo['bmHeight']), bmpstr, 'raw', 'BGRX', 0, 1) win32gui.DeleteObject(saveBitMap.GetHandle()) saveDC.DeleteDC() mfcDC.DeleteDC() win32gui.ReleaseDC(hwnd, hwndDC) return ocr_image(img, lang, enhance, engine) if __name__ == "__main__": r = ocr_screen((0, 0, 400, 100)) print(f"识别结果: {r['text']}") for line in r['lines']: print(f" 行: {line}") if 'details' in r: for d in r['details']: print(f" [{d['conf']:.3f}] {d['text']}")