--- layout: single title: "开源 OCR 引擎基准测试" date: 2023-10-09 08:00:00 +0800 categories: [容器与云原生, 操作系统] tags: [EasyOCR, PaddleOCR, OCR, Benchmark, faulthandler, WALL-E-AI] --- ## OCR 引擎 ### [EasyOCR](https://github.com/JaidedAI/EasyOCR) EasyOCR 支持 80+ 语言。 ``` Abaza = 'abq' Adyghe = 'ady' Afrikaans = 'af' Angika = 'ang' Arabic = 'ar' Assamese = 'as' Avar = 'ava' Azerbaijani = 'az' Belarusian = 'be' Bulgarian = 'bg' Bihari = 'bh' Bhojpuri = 'bho' Bengali = 'bn' Bosnian = 'bs' Simplified_Chinese = 'ch_sim' Traditional_Chinese = 'ch_tra' Chechen = 'che' Czech = 'cs' Welsh = 'cy' Danish = 'da' Dargwa = 'dar' German = 'de' English = 'en' Spanish = 'es' Estonian = 'et' Persian_Farsi = 'fa' French = 'fr' Irish = 'ga' Goan_Konkani = 'gom' Hindi = 'hi' Croatian = 'hr' Hungarian = 'hu' Indonesian = 'id' Ingush = 'inh' Icelandic = 'is' Italian = 'it' Japanese = 'ja' Kabardian = 'kbd' Kannada = 'kn' Korean = 'ko' Kurdish = 'ku' Latin = 'la' Lak = 'lbe' Lezghian = 'lez' Lithuanian = 'lt' Latvian = 'lv' Magahi = 'mah' Maithili = 'mai' Maori = 'mi' Mongolian = 'mn' Marathi = 'mr' Malay = 'ms' Maltese = 'mt' Nepali = 'ne' Newari = 'new' Dutch = 'nl' Norwegian = 'no' Occitan = 'oc' Pali = 'pi' Polish = 'pl' Portuguese = 'pt' Romanian = 'ro' Russian = 'ru' Serbian_cyrillic = 'rs_cyrillic' Serbian_latin = 'rs_latin' Nagpuri = 'sck' Slovak = 'sk' Slovenian = 'sl' Albanian = 'sq' Swedish = 'sv' Swahili = 'sw' Tamil = 'ta' Tabassaran = 'tab' Telugu = 'te' Thai = 'th' Tajik = 'tjk' Tagalog = 'tl' Turkish = 'tr' Uyghur = 'ug' Ukranian = 'uk' Urdu = 'ur' Uzbek = 'uz' Vietnamese = 'vi' ``` #### 安装 ```bash pip install torch==2.0.1 torchvision==0.15.2 -i https://download.pytorch.org/whl/cpu pip install easyocr ``` #### 代码示例 ```py import easyocr languages = ['ch_sim', 'en'] model = easyocr.Reader(languages) result = model.readtext(img_path) print(result) ``` ``` [([[28, 2], [372, 2], [372, 58], [28, 58]], '瓦力项目管理系统', 0.7662331484609081), ([[393, 11], [451, 11], [451, 47], [393, 47]], '首页', 0.9506149347953033), ([[460, 12], [562, 12], [562, 44], [460, 44]], '项目管理', 0.5817481294832926)] ``` ### [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR) PaddleOCR 支持中英文、英文、法语、德语、韩语、日语。 ``` Simplified_Chinese = 'ch' English = 'en' French = 'french' German = 'german' Korean = 'korean' Japanese = 'japan' ``` #### 安装 ```bash pip install numpy==1.23.5 PyMuPDF==1.21.1 # paddleocr 依赖 pip install paddlepaddle -i https://pypi.tuna.tsinghua.edu.cn/simple pip install paddleocr ``` #### 代码示例 ```py from paddleocr import PaddleOCR PADDLEOCR_MODULE_PATH = 'models/paddleocr' language = ['ch'] model = PaddleOCR(use_angle_cls=True, lang=language, det_limit_side_len=2000, det_model_dir=os.path.join(PADDLEOCR_MODULE_PATH, 'det'), rec_model_dir=os.path.join(PADDLEOCR_MODULE_PATH, 'rec'), cls_model_dir=os.path.join(PADDLEOCR_MODULE_PATH, 'cls')) result = model.ocr(img_path) print(result) ``` ``` [[[[36.0, 13.0], [361.0, 13.0], [361.0, 45.0], [36.0, 45.0]], ('瓦力项自管理系统', 0.9391518235206604)], [[[397.0, 15.0], [466.0, 15.0], [466.0, 41.0], [397.0, 41.0]], ('首页：', 0.8162643909454346)], [[[455.0, 16.0], [563.0, 16.0], [563.0, 38.0], [455.0, 38.0]], ('项目管理', 0.9921224117279053)]] ``` ### [Tesseract OCR](https://github.com/tesseract-ocr/tesseract) ## 效果图 ### EasyOCR ![](/images/2023/ocr/easyocr_result.png) ### PaddleOCR #### det_limit_side_len=960 (default) ![](/images/2023/ocr/paddleocr_result-det-limit-side-len-960.png) #### det_limit_side_len=2000 ![](/images/2023/ocr/paddleocr_result-det-limit-side-len-2000.png) **EasyOCR的效果更好** ## 镜像构建 ### EasyOCR #### CPU | Package | Version ✅ | Version ❌ | | --- | --- | --- | | easyocr | 1.7.1 | 1.7.1 | | torch | 2.0.1 | 2.1.0 | | torchvision | 0.15.2 | 0.16.0 | ❌ `torch` 升级到最新版本后，`easyocr` 会报错：`Segmentation fault`。 包依赖关系 #### EasyOCR | Package | Version | | ------------ | ------- | | easyocr | 1.7.1 | | torch | 2.0.1 | | torchvision | 0.15.2 | #### PaddleOCR | Package | Version | | ------------ | ------- | | paddleocr | 2.6.0 | | numpy | 1.23.5 | | paddlepaddle | 2.5.1 | | PyMuPDF | 1.21.1 | ```sh python app/aimodels/ocr_aimodel.pyc ``` ``` 🚗 Loading EasyOCR model... Neither CUDA nor MPS are available - defaulting to CPU. Note: This module is much faster with a GPU. Segmentation fault ``` ❌ 在 Apple M2 Max 上用 CPU 计算是通过容器的方式运行的 `amd64` 平台，出现错误：`Illegal instruction`。 ```sh docker run --rm -it -e MAX_WORKERS=1 wangjunjian/wall-e-ai:amd64 bash ``` ``` WARNING: The requested image's platform (linux/amd64) does not match the detected host platform (linux/arm64/v8) and no specific platform was requested root@daa9a1886e87:/WALL-E-AI# export PYTHONPATH=`pwd` root@daa9a1886e87:/WALL-E-AI# python app/aimodels/ocr_aimodel.pyc 🚗 Loading EasyOCR model... Neither CUDA nor MPS are available - defaulting to CPU. Note: This module is much faster with a GPU. [W NNPACK.cpp:64] Could not initialize NNPACK! Reason: Unsupported hardware. Illegal instruction ``` **虽然在 macOS 上运行出现错误，但是在 Linux amd64 平台上运行是正常的。** ### PaddleOCR #### CPU ❌ 在 Apple M2 Max 上用 CPU 计算是通过容器的方式运行的 `amd64` 平台，出现错误：`Illegal instruction`。 ```sh docker run --rm -it -p 9999:80 -e MAX_WORKERS=1 wangjunjian/wall-e-ai:amd64-paddleocr bash ``` ``` WARNING: The requested image's platform (linux/amd64) does not match the detected host platform (linux/arm64/v8) and no specific platform was requested root@d0e437177daf:/WALL-E-AI# export PYTHONPATH=`pwd` root@d0e437177daf:/WALL-E-AI# python app/aimodels/ocr_aimodel.pyc 🚗 Loading PaddleOCR model... Illegal instruction ``` **虽然在 macOS 上运行出现错误，但是在 Linux amd64 平台上运行是正常的。** ❌ 构建 `arm64` 平台的镜像，出现找不到 `paddlepaddle` 软件包的错误。 ## 速度测试 | | EasyOCR 1 (MS) | EasyOCR 2~ (MS) | PaddleOCR 1 (MS) | PaddleOCR 2~ (MS) | | --------------------------------- | -------------- | --------------- | ---------------- | ----------------- | | **CPU (Intel Xeon Silver 4216)** | 11015 | 6184（🚀 35%） | 11508 | 8135 | | **CPU (Apple M2 Max)** | 36398 | 31618 | 5334 | 3939 | | **MPS (Apple M2 Max)** | 10271 | 6792 | X | X | | **CUDA (NVIDIA Tesla T4)** | 18727 | 1194 | 17780 | 784（🚀 52%） | * 1 指的是第一次识别，需要加载模型，所以速度会慢。 * 2~ 指的是第二次识别的平均速度。 * X 表示不支持。 * PaddleOCR 参数 ```det_limit_side_len``` 设置为 2000，主要是为了提高大图的识别精度。 * EasyOCR 在 Apple M2 Max 上用 CPU 计算是通过容器的方式运行的 arm64 平台。 * EasyOCR 在 T4 上需要 ```3G``` 的显存。 * PaddleOCR 在 T4 上需要 ```2G``` 的显存。 ## 调试：[faulthandler — Dump the Python traceback](https://docs.python.org/3/library/faulthandler.html) 当出现下面的错误（主要是 C、C++ 库的问题）： * Segmentation fault * Illegal instruction 可以通过 `faulthandler` 定位到错误位置，在代码中加入下面的代码： ```py import faulthandler faulthandler.enable() ``` 代码实现如下： ```py import faulthandler faulthandler.enable() import os from app.config import * from app.aimodels.aimodel import AIModel from app.models.base_model import Object, Box class EasyOCR(AIModel): DEFAULT_LANGUAGE_CHINESE_ENGLISH = ['ch_sim', 'en'] def __init__(self, languages = DEFAULT_LANGUAGE_CHINESE_ENGLISH): super().__init__() self.languages = languages self._load_model() def _load_model(self): print(f'🚗 Loading {self.__class__.__name__} model...') import easyocr self.model = easyocr.Reader(self.languages) def load(self, languages = DEFAULT_LANGUAGE_CHINESE_ENGLISH): self.languages = languages self._load_model() def __call__(self, img_path): result = self.model.readtext(img_path) return result if __name__ == "__main__": img_path = 'static/docs/images/vision/image_match/crop_menuitem1.png' easyocr = EasyOCR() easyocr(img_path) ``` 运行代码后，出现下面的错误信息： ```bash Neither CUDA nor MPS are available - defaulting to CPU. Note: This module is much faster with a GPU. Fatal Python error: Segmentation fault Current thread 0x0000ffff976b7d40 (most recent call first): File "/usr/local/lib/python3.10/site-packages/torch/_ops.py", line 692 in __call__ File "/usr/local/lib/python3.10/site-packages/torch/ao/nn/quantized/dynamic/modules/rnn.py", line 119 in __init__ File "/usr/local/lib/python3.10/site-packages/torch/ao/nn/quantized/dynamic/modules/rnn.py", line 400 in __init__ File "/usr/local/lib/python3.10/site-packages/torch/ao/nn/quantized/dynamic/modules/rnn.py", line 291 in from_float File "/usr/local/lib/python3.10/site-packages/torch/ao/nn/quantized/dynamic/modules/rnn.py", line 495 in from_float File "/usr/local/lib/python3.10/site-packages/torch/ao/quantization/quantize.py", line 623 in swap_module File "/usr/local/lib/python3.10/site-packages/torch/ao/quantization/quantize.py", line 590 in _convert File "/usr/local/lib/python3.10/site-packages/torch/ao/quantization/quantize.py", line 588 in _convert File "/usr/local/lib/python3.10/site-packages/torch/ao/quantization/quantize.py", line 588 in _convert File "/usr/local/lib/python3.10/site-packages/torch/ao/quantization/quantize.py", line 550 in convert File "/usr/local/lib/python3.10/site-packages/torch/ao/quantization/quantize.py", line 465 in quantize_dynamic File "/usr/local/lib/python3.10/site-packages/easyocr/recognition.py", line 177 in get_recognizer File "/usr/local/lib/python3.10/site-packages/easyocr/easyocr.py", line 231 in __init__ File "/WALL-E-AI/app/aimodels/ocr_aimodel.py", line 25 in _load_model File "/WALL-E-AI/app/aimodels/ocr_aimodel.py", line 19 in __init__ File "/WALL-E-AI/app/aimodels/aimodel.py", line 11 in __call__ File "/WALL-E-AI/app/aimodels/ocr_aimodel.py", line 92 in Extension modules: PIL._imaging, numpy.core._multiarray_umath, numpy.core._multiarray_tests, numpy.linalg._umath_linalg, numpy.fft._pocketfft_internal, numpy.random._common, numpy.random.bit_generator, numpy.random._bounded_integers, numpy.random._mt19937, numpy.random.mtrand, numpy.random._philox, numpy.random._pcg64, numpy.random._sfc64, numpy.random._generator, torch._C, torch._C._fft, torch._C._linalg, torch._C._nested, torch._C._nn, torch._C._sparse, torch._C._special, PIL._imagingft, charset_normalizer.md, scipy._lib._ccallback_c, scipy.ndimage._nd_image, scipy.special._ufuncs_cxx, scipy.special._ufuncs, scipy.special._specfun, scipy.special._comb, scipy.linalg._fblas, scipy.linalg._flapack, scipy.linalg.cython_lapack, scipy.linalg._cythonized_array_utils, scipy.linalg._solve_toeplitz, scipy.linalg._decomp_lu_cython, scipy.linalg._matfuncs_sqrtm_triu, scipy.linalg.cython_blas, scipy.linalg._matfuncs_expm, scipy.linalg._decomp_update, scipy.sparse._sparsetools, _csparsetools, scipy.sparse._csparsetools, scipy.sparse.linalg._isolve._iterative, scipy.sparse.linalg._dsolve._superlu, scipy.sparse.linalg._eigen.arpack._arpack, scipy.sparse.csgraph._tools, scipy.sparse.csgraph._shortest_path, scipy.sparse.csgraph._traversal, scipy.sparse.csgraph._min_spanning_tree, scipy.sparse.csgraph._flow, scipy.sparse.csgraph._matching, scipy.sparse.csgraph._reordering, scipy.linalg._flinalg, scipy.special._ellip_harm_2, _ni_label, scipy.ndimage._ni_label, skimage._shared.geometry, yaml._yaml (total: 58) Segmentation fault ``` ## 参考资料 * [PyTorch Install](https://pytorch.org/get-started/locally/) * [pypi PaddleOCR](https://pypi.org/project/paddleocr/) * [PaddlePaddle Install](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/install/pip/linux-pip.html) * [Dockerfile reference](https://docs.docker.com/engine/reference/builder/) * [Docker python](https://hub.docker.com/_/python/) * [Docker PyTorch](https://hub.docker.com/r/pytorch/) * [Docker 中 arg 和 env 的区别](https://blog.justwe.site/post/docker-arg-env/) * [中文开源字体集](https://drxie.github.io/OSFCC/) * [PaddleOCR学习 - det_limit_side_len](https://codeantenna.com/a/2q60ScwEBo) * [JaidedAI/EasyOCR](https://github.com/JaidedAI/EasyOCR) * [EasyOCR vs Tesseract vs Amazon Textract: an OCR engine comparison](https://francescopochetti.com/easyocr-vs-tesseract-vs-amazon-textract-an-ocr-engine-comparison/) * [Training and Deploying a fully Dockerized License Plate Recognition app with IceVision, Amazon Textract and FastAPI](https://francescopochetti.com/training-and-deploying-a-fully-dockerized-license-plate-recognition-app-with-icevision-amazon-textract-and-fastapi/) * [Query on image resolution used for detection and recognition #1078 default=2560](https://github.com/JaidedAI/EasyOCR/issues/1078)