{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "## 1. PP-OCRv3模型简介\n", "\n", "PP-OCRv3在PP-OCRv2的基础上进一步升级。整体的框架图保持了与PP-OCRv2相同的pipeline，针对检测模型和识别模型进行了优化。其中，检测模块仍基于DB算法优化，而识别模块不再采用CRNN，换成了IJCAI 2022最新收录的文本识别算法SVTR，并对其进行产业适配。PP-OCRv3系统框图如下所示（粉色框中为PP-OCRv3新增策略）：\n", "\n", "
\n", "\n", "
\n", "\n", "\n", "从算法改进思路上看，分别针对检测和识别模型，进行了共9个方面的改进：\n", "\n", "- 检测模块：\n", " - LK-PAN：大感受野的PAN结构；\n", " - DML：教师模型互学习策略；\n", " - RSE-FPN：残差注意力机制的FPN结构；\n", "\n", "\n", "- 识别模块：\n", " - SVTR_LCNet：轻量级文本识别网络；\n", " - GTC：Attention指导CTC训练策略；\n", " - TextConAug：挖掘文字上下文信息的数据增广策略；\n", " - TextRotNet：自监督的预训练模型；\n", " - UDML：联合互学习策略；\n", " - UIM：无标注数据挖掘方案。\n", "\n", "从效果上看，速度可比情况下，多种场景精度均有大幅提升：\n", "- 中文场景，相对于PP-OCRv2中文模型提升超5%；\n", "- 英文数字场景，相比于PP-OCRv2英文模型提升11%；\n", "- 多语言场景，优化80+语种识别效果，平均准确率提升超5%。\n", "\n", "\n", "更详细的优化细节可参考技术报告：https://arxiv.org/abs/2206.03001 。\n", "\n", "更多关于PaddleOCR的内容，可以点击 https://github.com/PaddlePaddle/PaddleOCR 进行了解。\n", "\n", "\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 2. 模型效果\n", "\n", "PP-OCRv3的效果如下：\n", "\n", "
\n", "\n", "
\n", "
\n", "\n", "
\n", "\n", "\n", "\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 3. 模型如何使用\n", "\n", "### 3.1 模型推理\n", "* 安装PaddleOCR whl包" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false, "jupyter": { "outputs_hidden": false }, "scrolled": true, "tags": [] }, "outputs": [], "source": [ "! pip install paddleocr --user" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "* 快速体验" ] }, { "cell_type": "code", "execution_count": 1, "metadata": { "scrolled": true, "tags": [] }, "outputs": [], "source": [ "# 命令行使用\n", "! wget https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/imgs/11.jpg\n", "! paddleocr --image_dir 11.jpg --use_angle_cls true" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "运行完成后，会在终端输出如下结果：\n", "```log\n", "[[[28.0, 37.0], [302.0, 39.0], [302.0, 72.0], [27.0, 70.0]], ('纯臻营养护发素', 0.96588134765625)]\n", "[[[26.0, 81.0], [172.0, 83.0], [172.0, 104.0], [25.0, 101.0]], ('产品信息/参数', 0.9113278985023499)]\n", "[[[28.0, 115.0], [330.0, 115.0], [330.0, 132.0], [28.0, 132.0]], ('（45元/每公斤，100公斤起订）', 0.8843421936035156)]\n", "......\n", "```\n", "\n", "\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 3.2 模型训练\n", "PP-OCR系统由文本检测模型、方向分类器和文本识别模型构成，三个模型训练教程可参考如下文档:\n", "1. 文本检测模型：[文本检测训练教程](https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.6/doc/doc_ch/detection.md)\n", "1. 方向分类器: [方向分类器训练教程](https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.6/doc/doc_ch/angle_class.md)\n", "1. 文本识别模型：[文本识别训练教程](https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.6/doc/doc_ch/recognition.md)\n", "\n", "模型训练完成后，可以通过指定模型路径的方式串联使用\n", "命令参考如下：\n", "```python\n", "paddleocr --image_dir 11.jpg --use_angle_cls true --det_model_dir=/path/to/det_inference_model --cls_model_dir=/path/to/cls_inference_model --rec_model_dir=/path/to/rec_inference_model\n", "```" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 4. 原理\n", "\n", "优化思路具体如下\n", "\n", "1. 检测模型优化\n", "- LK-PAN：大感受野的PAN结构。\n", " \n", "LK-PAN (Large Kernel PAN) 是一个具有更大感受野的轻量级PAN结构，核心是将PAN结构的path augmentation中卷积核从3*3改为9*9。通过增大卷积核，提升特征图每个位置覆盖的感受野，更容易检测大字体的文字以及极端长宽比的文字。使用LK-PAN结构，可以将教师模型的hmean从83.2%提升到85.0%。\n", "
\n", " \n", "
\n", "\n", "- DML：教师模型互学习策略\n", "\n", "[DML](https://arxiv.org/abs/1706.00384) （Deep Mutual Learning）互学习蒸馏方法，如下图所示，通过两个结构相同的模型互相学习，可以有效提升文本检测模型的精度。教师模型采用DML策略，hmean从85%提升到86%。将PP-OCRv2中CML的教师模型更新为上述更高精度的教师模型，学生模型的hmean可以进一步从83.2%提升到84.3%。\n", "
\n", " \n", "
\n", "\n", "- RSE-FPN：残差注意力机制的FPN结构\n", "\n", "RSE-FPN（Residual Squeeze-and-Excitation FPN）如下图所示，引入残差结构和通道注意力结构，将FPN中的卷积层更换为通道注意力结构的RSEConv层，进一步提升特征图的表征能力。考虑到PP-OCRv2的检测模型中FPN通道数非常小，仅为96，如果直接用SEblock代替FPN中卷积会导致某些通道的特征被抑制，精度会下降。RSEConv引入残差结构会缓解上述问题，提升文本检测效果。进一步将PP-OCRv2中CML的学生模型的FPN结构更新为RSE-FPN，学生模型的hmean可以进一步从84.3%提升到85.4%。\n", "\n", "
\n", "\n", "
\n", "\n", "1. 识别模型优化\n", "- SVTR_LCNet：轻量级文本识别网络\n", "\n", "SVTR_LCNet是针对文本识别任务，将基于Transformer的SVTR网络和轻量级CNN网络PP-LCNet 融合的一种轻量级文本识别网络。使用该网络，预测速度优于PP-OCRv2的识别模型20%，但是由于没有采用蒸馏策略，该识别模型效果略差。此外，进一步将输入图片规范化高度从32提升到48，预测速度稍微变慢，但是模型效果大幅提升，识别准确率达到73.98%（+2.08%），接近PP-OCRv2采用蒸馏策略的识别模型效果。\n", "\n", "- GTC：Attention指导CTC训练策略\n", " \n", "[GTC](https://arxiv.org/pdf/2002.01276.pdf)（Guided Training of CTC），利用Attention模块CTC训练，融合多种文本特征的表达，是一种有效的提升文本识别的策略。使用该策略，预测时完全去除 Attention 模块，在推理阶段不增加任何耗时，识别模型的准确率进一步提升到75.8%（+1.82%）。训练流程如下所示：\n", "\n", "
\n", "\n", "
\n", "\n", "- TextConAug：挖掘文字上下文信息的数据增广策略\n", "\n", "TextConAug是一种挖掘文字上下文信息的数据增广策略，主要思想来源于论文[ConCLR](https://www.cse.cuhk.edu.hk/~byu/papers/C139-AAAI2022-ConCLR.pdf)，作者提出ConAug数据增广，在一个batch内对2张不同的图像进行联结，组成新的图像并进行自监督对比学习。PP-OCRv3将此方法应用到有监督的学习任务中，设计了TextConAug数据增强方法，可以丰富训练数据上下文信息，提升训练数据多样性。使用该策略，识别模型的准确率进一步提升到76.3%（+0.5%）。TextConAug示意图如下所示：\n", "\n", "
\n", "\n", "
\n", "\n", "- TextRotNet：自监督的预训练模型\n", "\n", "TextRotNet是使用大量无标注的文本行数据，通过自监督方式训练的预训练模型，参考于论文[STR-Fewer-Labels](https://github.com/ku21fan/STR-Fewer-Labels)。该模型可以初始化SVTR_LCNet的初始权重，从而帮助文本识别模型收敛到更佳位置。使用该策略，识别模型的准确率进一步提升到76.9%（+0.6%）。TextRotNet训练流程如下图所示：\n", "\n", "
\n", "\n", "
\n", "\n", "- UDML：联合互学习策略\n", "\n", "UDML（Unified-Deep Mutual Learning）联合互学习是PP-OCRv2中就采用的对于文本识别非常有效的提升模型效果的策略。在PP-OCRv3中，针对两个不同的SVTR_LCNet和Attention结构，对他们之间的PP-LCNet的特征图、SVTR模块的输出和Attention模块的输出同时进行监督训练。使用该策略，识别模型的准确率进一步提升到78.4%（+1.5%）。\n", "\n", "- UDML：联合互学习策略\n", "\n", "UIM（Unlabeled Images Mining）是一种非常简单的无标注数据挖掘方案。核心思想是利用高精度的文本识别大模型对无标注数据进行预测，获取伪标签，并且选择预测置信度高的样本作为训练数据，用于训练小模型。使用该策略，识别模型的准确率进一步提升到79.4%（+1%）。\n", "\n", "
\n", "\n", "
" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 5. 注意事项\n", "\n", "PP-OCR系列模型训练过程中均使用通用数据，如在实际场景中表现不满意，可标注少量数据进行finetune。" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## 6. 相关论文以及引用信息\n", "```\n", "@article{li2022pp,\n", " title={PP-OCRv3: More Attempts for the Improvement of Ultra Lightweight OCR System},\n", " author={Li, Chenxia and Liu, Weiwei and Guo, Ruoyu and Yin, Xiaoting and Jiang, Kaitao and Du, Yongkun and Du, Yuning and Zhu, Lingfeng and Lai, Baohua and Hu, Xiaoguang and others},\n", " journal={arXiv preprint arXiv:2206.03001},\n", " year={2022}\n", "}\n", "```\n" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3.8.13 ('py38')", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.8.13" }, "vscode": { "interpreter": { "hash": "58fd1890da6594cebec461cf98c6cb9764024814357f166387d10d267624ecd6" } } }, "nbformat": 4, "nbformat_minor": 4 }