Skip to content

GitLab

Commit c89bc397 authored by andyjpaddle's avatar andyjpaddle
Browse files

Merge branch 'release/2.5' of https://github.com/PaddlePaddle/PaddleOCR into release/2.5

parents 2274364b 61b03628
Expand all Show whitespace changes
Inline Side-by-side
Showing with 988 additions and 26 deletions
PPOCRLabel/README.md
View file @ c89bc397
......@@ -8,7 +8,7 @@ PPOCRLabelv2 is a semi-automatic graphic annotation tool suitable for OCR field,
| :-------------------------------------------------: | :--------------------------------------------: |
| <img src="./data/gif/steps_en.gif" width="80%"/> | <img src="./data/gif/table.gif" width="100%"/> |
| **irregular text annotation** | **key information annotation** |
| <img src="./data/gif/multi-point.gif" width="80%"/> | <img src="./data/gif/kie.gif" width="300%"/> |
| <img src="./data/gif/multi-point.gif" width="80%"/> | <img src="./data/gif/kie.gif" width="100%"/> |
### Recent Update
......
PPOCRLabel/README_ch.md
View file @ c89bc397
......@@ -8,7 +8,7 @@ PPOCRLabel是一款适用于OCR领域的半自动化图形标注工具，内置P
| :---------------------------------------------------: | :----------------------------------------------: |
| <img src="./data/gif/steps_en.gif" width="80%"/> | <img src="./data/gif/table.gif" width="100%"/> |
| **不规则文本标注** | **关键信息标注** |
| <img src="./data/gif/multi-point.gif" width="80%"/> | <img src="./data/gif/kie.gif" width="300%"/> |
| <img src="./data/gif/multi-point.gif" width="80%"/> | <img src="./data/gif/kie.gif" width="100%"/> |
#### 近期更新
......
README.md
View file @ c89bc397
......@@ -26,6 +26,8 @@ PaddleOCR aims to create multilingual, awesome, leading, and practical OCR tools
</div>
## Recent updates
- **🔥2022.7 Release [OCR scene application collection](./applications/README_en.md)**
- PaddleOCR scene application covers general, manufacturing, finance, transportation industry of the main OCR vertical applications, including digital tube, LCD screen character, license plate, high-precision SVTR model, etc. **7 vertical models**.
- **🔥2022.5.9 Release PaddleOCR [release/2.5](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.5)**
- Release [PP-OCRv3](./doc/doc_en/ppocr_introduction_en.md#pp-ocrv3): With comparable speed, the effect of Chinese scene is further improved by 5% compared with PP-OCRv2, the effect of English scene is improved by 11%, and the average recognition accuracy of 80 language multilingual models is improved by more than 5%.
- Release [PPOCRLabelv2](./PPOCRLabel): Add the annotation function for table recognition task, key information extraction task and irregular text image.
......@@ -37,7 +39,6 @@ PaddleOCR aims to create multilingual, awesome, leading, and practical OCR tools
- Release [PP-OCRv2](./doc/doc_en/ppocr_introduction_en.md#pp-ocrv2). The inference speed of PP-OCRv2 is 220% higher than that of PP-OCR server in CPU device. The F-score of PP-OCRv2 is 7% higher than that of PP-OCR mobile.
- 2021.8.3 Release PaddleOCR [release/2.2](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.2)
- Release a new structured documents analysis toolkit, i.e., [PP-Structure](./ppstructure/README.md), support layout analysis and table recognition (One-key to export chart images to Excel files).
- [more](./doc/doc_en/update_en.md)
......@@ -47,14 +48,15 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel
![](./doc/features_en.png)
> It is recommended to start with the “quick experience” in the document tutorial
> It is recommended to start with the “quick start” in the document tutorial
## Quick Experience
- One line of code quick use: [Quick Start](./doc/doc_en/quickstart_en.md)
- Web online experience for the ultra-lightweight OCR: [Online Experience](https://www.paddlepaddle.org.cn/hub/scene/ocr)
- Mobile DEMO experience (based on EasyEdge and Paddle-Lite, supports iOS and Android systems): [Sign in to the website to obtain the QR code for installing the App](https://ai.baidu.com/easyedge/app/openSource?from=paddlelite)
- One line of code quick use: [Quick Start](./doc/doc_en/quickstart_en.md)
<a name="book"></a>
......@@ -63,14 +65,16 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel
<a name="Community"></a>
## Community
## Community👬
- **Join us**👬: Scan the QR code below with your Wechat, you can join the official technical discussion group. Looking forward to your participation.
- For international developers, we regard [PaddleOCR Discussions](https://github.com/PaddlePaddle/PaddleOCR/discussions) as our international community platform. All ideas and questions can be discussed here in English.
- For Chinese develops, Scan the QR code below with your Wechat, you can join the official technical discussion group. For richer community content, please refer to [中文README](README_ch.md), looking forward to your participation.
<div align="center">
<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "200" height = "200" />
<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "150" height = "150" />
</div>
<a name="Supported-Chinese-model-list"></a>
## PP-OCR Series Model List（Update on September 8th）
......
README_ch.md
View file @ c89bc397
......@@ -27,16 +27,8 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
## 近期更新
- **🔥2022.5.25~26 OCR产业应用两日直播课**
- 25日：车牌识别产业应用实战（[AI Studio项目链接](https://aistudio.baidu.com/aistudio/projectdetail/3919091?contributionType=1)）
- 26日：一招搞定工业常见数码管、PCB字符识别（AI Studio项目链接：[数码管识别](https://aistudio.baidu.com/aistudio/projectdetail/4049044?contributionType=1)，[PCB字符识别](https://aistudio.baidu.com/aistudio/projectdetail/4008973)）
扫描下方二维码填写问卷后进入群聊，获取直播链接！
<div align="center">
<img src="https://user-images.githubusercontent.com/50011306/170023861-38814d84-b35a-4102-94d9-28482f9a39f8.png" width = "150" height = "150" />
</div>
- **🔥2022.7 发布[OCR场景应用集合](./applications)**
- 发布OCR场景应用集合，包含数码管、液晶屏、车牌、高精度SVTR模型等**7个垂类模型**，覆盖通用，制造、金融、交通行业的主要OCR垂类应用。
- **🔥2022.5.9 发布PaddleOCR [release/2.5](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.5)**
- 发布[PP-OCRv3](./doc/doc_ch/ppocr_introduction.md#pp-ocrv3)，速度可比情况下，中文场景效果相比于PP-OCRv2再提升5%，英文场景提升11%，80语种多语言模型平均识别准确率提升5%以上；
......@@ -75,19 +67,19 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
<a name="开源社区"></a>
## 开源社区
- **项目合作📑：** 如果您是企业开发者且有明确的OCR垂类应用需求，填写[问卷](https://paddle.wjx.cn/vj/QwF7GKw.aspx)后可免费与官方团队展开不同层次的合作。
- **加入社区👬：** 微信扫描二维码并填写问卷之后，加入交流群领取福利
- **获取PaddleOCR最新发版解说《OCR超强技术详解与产业应用实战》系列直播课回放链接**
- **10G重磅OCR学习大礼包：**《动手学OCR》电子书，配套讲解视频和notebook项目；66篇OCR相关顶会前沿论文打包放送，包括CVPR、AAAI、IJCAI、ICCV等；PaddleOCR历次发版直播课视频；OCR社区优秀开发者项目分享视频。
- **社区贡献**🏅️：[社区贡献](./doc/doc_ch/thirdparty.md)文档中包含了社区用户**使用PaddleOCR开发的各种工具、应用**以及**为PaddleOCR贡献的功能、优化的文档与代码**等，是官方为社区开发者打造的荣誉墙，也是帮助优质项目宣传的广播站。
- **社区项目**🏅️：[社区项目](./doc/doc_ch/thirdparty.md)文档中包含了社区用户**使用PaddleOCR开发的各种工具、应用**以及**为PaddleOCR贡献的功能、优化的文档与代码**等，是官方为社区开发者打造的荣誉墙，也是帮助优质项目宣传的广播站。
- **社区常规赛**🎁：社区常规赛是面向OCR开发者的积分赛事，覆盖文档、代码、模型和应用四大类型，以季度为单位评选并发放奖励，赛题详情与报名方法可参考[链接](https://github.com/PaddlePaddle/PaddleOCR/issues/4982)。
<div align="center">
<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "200" height = "200" />
<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "150" height = "150" />
</div>
<a name="模型下载"></a>
## PP-OCR系列模型列表（更新中）
......@@ -95,14 +87,21 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
| ------------------------------------- | ----------------------- | --------------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ |
| 中英文超轻量PP-OCRv3模型（16.2M） | ch_PP-OCRv3_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar) |
| 英文超轻量PP-OCRv3模型（13.4M） | en_PP-OCRv3_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_distill_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_train.tar) |
| 中英文超轻量PP-OCRv2模型（13.0M） | ch_PP-OCRv2_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar) |
| 中英文超轻量PP-OCR mobile模型（9.4M） | ch_ppocr_mobile_v2.0_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_pre.tar) |
| 中英文通用PP-OCR server模型（143.4M） | ch_ppocr_server_v2.0_xx | 服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_pre.tar) |
更多模型下载（包括多语言），可以参考[PP-OCR 系列模型下载](./doc/doc_ch/models_list.md)，文档分析相关模型参考[PP-Structure 系列模型下载](./ppstructure/docs/models_list.md)
- 超轻量OCR系列更多模型下载（包括多语言），可以参考[PP-OCR系列模型下载](./doc/doc_ch/models_list.md)，文档分析相关模型参考[PP-Structure系列模型下载](./ppstructure/docs/models_list.md)
### PaddleOCR场景应用模型
| 行业 | 类别 | 亮点 | 文档说明 | 模型下载 |
| ---- | ------------ | ---------------------------------- | ------------------------------------------------------------ | --------------------------------------------- |
| 制造 | 数码管识别 | 数码管数据合成、漏识别调优 | [光功率计数码管字符识别](./applications/光功率计数码管字符识别/光功率计数码管字符识别.md) | [下载链接](./applications/README.md#模型下载) |
| 金融 | 通用表单识别 | 多模态通用表单结构化提取 | [多模态表单识别](./applications/多模态表单识别.md) | [下载链接](./applications/README.md#模型下载) |
| 交通 | 车牌识别 | 多角度图像处理、轻量模型、端侧部署 | [轻量级车牌识别](./applications/轻量级车牌识别.md) | [下载链接](./applications/README.md#模型下载) |
- 更多制造、金融、交通行业的主要OCR垂类应用模型（如电表、液晶屏、高精度SVTR模型等），可参考[场景应用模型下载](./applications)
<a name="文档教程"></a>
## 文档教程
- [运行环境准备](./doc/doc_ch/environment.md)
......
applications/PCB字符识别/PCB字符识别.md 0 → 100644
View file @ c89bc397
This diff is collapsed.
applications/PCB字符识别/gen_data/corpus/text.txt 0 → 100644
View file @ c89bc397
5ZQ
I4UL
PWL
SNOG
ZL02
1C30
O3H
YHRS
N03S
1U5Y
JTK
EN4F
YKJ
DWNH
R42W
X0V
4OF5
08AM
Y93S
GWE2
0KR
9U2A
DBQ
Y6J
ROZ
K06
KIEY
NZQJ
UN1B
6X4
\ No newline at end of file
applications/PCB字符识别/gen_data/gen.py 0 → 100644
View file @ c89bc397
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This code is refer from:
https://github.com/zcswdt/Color_OCR_image_generator
"""
import os
import random
from PIL import Image, ImageDraw, ImageFont
import json
import argparse
def get_char_lines(txt_root_path):
"""
desc:get corpus line
"""
txt_files = os.listdir(txt_root_path)
char_lines = []
for txt in txt_files:
f = open(os.path.join(txt_root_path, txt), mode='r', encoding='utf-8')
lines = f.readlines()
f.close()
for line in lines:
char_lines.append(line.strip())
return char_lines
def get_horizontal_text_picture(image_file, chars, fonts_list, cf):
"""
desc:gen horizontal text picture
"""
img = Image.open(image_file)
if img.mode != 'RGB':
img = img.convert('RGB')
img_w, img_h = img.size
# random choice font
font_path = random.choice(fonts_list)
# random choice font size
font_size = random.randint(cf.font_min_size, cf.font_max_size)
font = ImageFont.truetype(font_path, font_size)
ch_w = []
ch_h = []
for ch in chars:
wt, ht = font.getsize(ch)
ch_w.append(wt)
ch_h.append(ht)
f_w = sum(ch_w)
f_h = max(ch_h)
# add space
char_space_width = max(ch_w)
f_w += (char_space_width * (len(chars) - 1))
x1 = random.randint(0, img_w - f_w)
y1 = random.randint(0, img_h - f_h)
x2 = x1 + f_w
y2 = y1 + f_h
crop_y1 = y1
crop_x1 = x1
crop_y2 = y2
crop_x2 = x2
best_color = (0, 0, 0)
draw = ImageDraw.Draw(img)
for i, ch in enumerate(chars):
draw.text((x1, y1), ch, best_color, font=font)
x1 += (ch_w[i] + char_space_width)
crop_img = img.crop((crop_x1, crop_y1, crop_x2, crop_y2))
return crop_img, chars
def get_vertical_text_picture(image_file, chars, fonts_list, cf):
"""
desc:gen vertical text picture
"""
img = Image.open(image_file)
if img.mode != 'RGB':
img = img.convert('RGB')
img_w, img_h = img.size
# random choice font
font_path = random.choice(fonts_list)
# random choice font size
font_size = random.randint(cf.font_min_size, cf.font_max_size)
font = ImageFont.truetype(font_path, font_size)
ch_w = []
ch_h = []
for ch in chars:
wt, ht = font.getsize(ch)
ch_w.append(wt)
ch_h.append(ht)
f_w = max(ch_w)
f_h = sum(ch_h)
x1 = random.randint(0, img_w - f_w)
y1 = random.randint(0, img_h - f_h)
x2 = x1 + f_w
y2 = y1 + f_h
crop_y1 = y1
crop_x1 = x1
crop_y2 = y2
crop_x2 = x2
best_color = (0, 0, 0)
draw = ImageDraw.Draw(img)
i = 0
for ch in chars:
draw.text((x1, y1), ch, best_color, font=font)
y1 = y1 + ch_h[i]
i = i + 1
crop_img = img.crop((crop_x1, crop_y1, crop_x2, crop_y2))
crop_img = crop_img.transpose(Image.ROTATE_90)
return crop_img, chars
def get_fonts(fonts_path):
"""
desc: get all fonts
"""
font_files = os.listdir(fonts_path)
fonts_list=[]
for font_file in font_files:
font_path=os.path.join(fonts_path, font_file)
fonts_list.append(font_path)
return fonts_list
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--num_img', type=int, default=30, help="Number of images to generate")
parser.add_argument('--font_min_size', type=int, default=11)
parser.add_argument('--font_max_size', type=int, default=12,
help="Help adjust the size of the generated text and the size of the picture")
parser.add_argument('--bg_path', type=str, default='./background',
help='The generated text pictures will be pasted onto the pictures of this folder')
parser.add_argument('--det_bg_path', type=str, default='./det_background',
help='The generated text pictures will use the pictures of this folder as the background')
parser.add_argument('--fonts_path', type=str, default='../../StyleText/fonts',
help='The font used to generate the picture')
parser.add_argument('--corpus_path', type=str, default='./corpus',
help='The corpus used to generate the text picture')
parser.add_argument('--output_dir', type=str, default='./output/', help='Images save dir')
cf = parser.parse_args()
# save path
if not os.path.exists(cf.output_dir):
os.mkdir(cf.output_dir)
# get corpus
txt_root_path = cf.corpus_path
char_lines = get_char_lines(txt_root_path=txt_root_path)
# get all fonts
fonts_path = cf.fonts_path
fonts_list = get_fonts(fonts_path)
# rec bg
img_root_path = cf.bg_path
imnames=os.listdir(img_root_path)
# det bg
det_bg_path = cf.det_bg_path
bg_pics = os.listdir(det_bg_path)
# OCR det files
det_val_file = open(cf.output_dir + 'det_gt_val.txt', 'w', encoding='utf-8')
det_train_file = open(cf.output_dir + 'det_gt_train.txt', 'w', encoding='utf-8')
# det imgs
det_save_dir = 'imgs/'
if not os.path.exists(cf.output_dir + det_save_dir):
os.mkdir(cf.output_dir + det_save_dir)
det_val_save_dir = 'imgs_val/'
if not os.path.exists(cf.output_dir + det_val_save_dir):
os.mkdir(cf.output_dir + det_val_save_dir)
# OCR rec files
rec_val_file = open(cf.output_dir + 'rec_gt_val.txt', 'w', encoding='utf-8')
rec_train_file = open(cf.output_dir + 'rec_gt_train.txt', 'w', encoding='utf-8')
# rec imgs
rec_save_dir = 'rec_imgs/'
if not os.path.exists(cf.output_dir + rec_save_dir):
os.mkdir(cf.output_dir + rec_save_dir)
rec_val_save_dir = 'rec_imgs_val/'
if not os.path.exists(cf.output_dir + rec_val_save_dir):
os.mkdir(cf.output_dir + rec_val_save_dir)
val_ratio = cf.num_img * 0.2 # val dataset ratio
print('start generating...')
for i in range(0, cf.num_img):
imname = random.choice(imnames)
img_path = os.path.join(img_root_path, imname)
rnd = random.random()
# gen horizontal text picture
if rnd < 0.5:
gen_img, chars = get_horizontal_text_picture(img_path, char_lines[i], fonts_list, cf)
ori_w, ori_h = gen_img.size
gen_img = gen_img.crop((0, 3, ori_w, ori_h))
# gen vertical text picture
else:
gen_img, chars = get_vertical_text_picture(img_path, char_lines[i], fonts_list, cf)
ori_w, ori_h = gen_img.size
gen_img = gen_img.crop((3, 0, ori_w, ori_h))
ori_w, ori_h = gen_img.size
# rec imgs
save_img_name = str(i).zfill(4) + '.jpg'
if i < val_ratio:
save_dir = os.path.join(rec_val_save_dir, save_img_name)
line = save_dir + '\t' + char_lines[i] + '\n'
rec_val_file.write(line)
else:
save_dir = os.path.join(rec_save_dir, save_img_name)
line = save_dir + '\t' + char_lines[i] + '\n'
rec_train_file.write(line)
gen_img.save(cf.output_dir + save_dir, quality = 95, subsampling=0)
# det img
# random choice bg
bg_pic = random.sample(bg_pics, 1)[0]
det_img = Image.open(os.path.join(det_bg_path, bg_pic))
# the PCB position is fixed, modify it according to your own scenario
if bg_pic == '1.png':
x1 = 38
y1 = 3
else:
x1 = 34
y1 = 1
det_img.paste(gen_img, (x1, y1))
# text pos
chars_pos = [[x1, y1], [x1 + ori_w, y1], [x1 + ori_w, y1 + ori_h], [x1, y1 + ori_h]]
label = [{"transcription":char_lines[i], "points":chars_pos}]
if i < val_ratio:
save_dir = os.path.join(det_val_save_dir, save_img_name)
det_val_file.write(save_dir + '\t' + json.dumps(
label, ensure_ascii=False) + '\n')
else:
save_dir = os.path.join(det_save_dir, save_img_name)
det_train_file.write(save_dir + '\t' + json.dumps(
label, ensure_ascii=False) + '\n')
det_img.save(cf.output_dir + save_dir, quality = 95, subsampling=0)
applications/README.md 0 → 100644
View file @ c89bc397
[English](README_en.md) | 简体中文
# 场景应用
PaddleOCR场景应用覆盖通用，制造、金融、交通行业的主要OCR垂类应用，在PP-OCR、PP-Structure的通用能力基础之上，以notebook的形式展示利用场景数据微调、模型优化方法、数据增广等内容，为开发者快速落地OCR应用提供示范与启发。
- [教程文档](#1)
- [通用](#11)
- [制造](#12)
- [金融](#13)
- [交通](#14)
- [模型下载](#2)
<a name="1"></a>
## 教程文档
<a name="11"></a>
### 通用
| 类别 | 亮点 | 模型下载 | 教程 |
| ---------------------- | ------------ | -------------- | --------------------------------------- |
| 高精度中文识别模型SVTR | 比PP-OCRv3识别模型精度高3%，可用于数据挖掘或对预测效率要求不高的场景。| [模型下载](#2) | [中文](./高精度中文识别模型.md)/English |
| 手写体识别 | 新增字形支持 | | |
<a name="12"></a>
### 制造
| 类别 | 亮点 | 模型下载 | 教程 | 示例图 |
| -------------- | ------------------------------ | -------------- | ------------------------------------------------------------ | ------------------------------------------------------------ |
| 数码管识别 | 数码管数据合成、漏识别调优 | [模型下载](#2) | [中文](./光功率计数码管字符识别/光功率计数码管字符识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/7d5774a273f84efba5b9ce7fd3f86e9ef24b6473e046444db69fa3ca20ac0986" width = "200" height = "100" /> |
| 液晶屏读数识别 | 检测模型蒸馏、Serving部署 | [模型下载](#2) | [中文](./液晶屏读数识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/901ab741cb46441ebec510b37e63b9d8d1b7c95f63cc4e5e8757f35179ae6373" width = "200" height = "100" /> |
| 包装生产日期 | 点阵字符合成、过曝过暗文字识别 | [模型下载](#2) | [中文](./包装生产日期识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/d9e0533cc1df47ffa3bbe99de9e42639a3ebfa5bce834bafb1ca4574bf9db684" width = "200" height = "100" /> |
| PCB文字识别 | 小尺寸文本检测与识别 | [模型下载](#2) | [中文](./PCB字符识别/PCB字符识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/95d8e95bf1ab476987f2519c0f8f0c60a0cdc2c444804ed6ab08f2f7ab054880" width = "200" height = "100" /> |
| 电表识别 | 大分辨率图像检测调优 | [模型下载](#2) | | |
| 液晶屏缺陷检测 | 非文字字符识别 | | | |
<a name="13"></a>
### 金融
| 类别 | 亮点 | 模型下载 | 教程 | 示例图 |
| -------------- | ------------------------ | -------------- | ----------------------------------- | ------------------------------------------------------------ |
| 表单VQA | 多模态通用表单结构化提取 | [模型下载](#2) | [中文](./多模态表单识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/a3b25766f3074d2facdf88d4a60fc76612f51992fd124cf5bd846b213130665b" width = "200" height = "200" /> |
| 增值税发票 | 尽请期待 | | | |
| 印章检测与识别 | 端到端弯曲文本识别 | | | |
| 通用卡证识别 | 通用结构化提取 | | | |
| 身份证识别 | 结构化提取、图像阴影 | | | |
| 合同比对 | 密集文本检测、NLP串联 | | | |
<a name="14"></a>
### 交通
| 类别 | 亮点 | 模型下载 | 教程 | 示例图 |
| ----------------- | ------------------------------ | -------------- | ----------------------------------- | ------------------------------------------------------------ |
| 车牌识别 | 多角度图像、轻量模型、端侧部署 | [模型下载](#2) | [中文](./轻量级车牌识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/76b6a0939c2c4cf49039b6563c4b28e241e11285d7464e799e81c58c0f7707a7" width = "200" height = "100" /> |
| 驾驶证/行驶证识别 | 尽请期待 | | | |
| 快递单识别 | 尽请期待 | | | |
<a name="2"></a>
## 模型下载
如需下载上述场景中已经训练好的垂类模型，可以扫描下方二维码，关注公众号填写问卷后，加入PaddleOCR官方交流群获取20G OCR学习大礼包（内含《动手学OCR》电子书、课程回放视频、前沿论文等重磅资料）
<div align="center">
<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e" width = "150" height = "150" />
</div>
如果您是企业开发者且未在上述场景中找到合适的方案，可以填写[OCR应用合作调研问卷](https://paddle.wjx.cn/vj/QwF7GKw.aspx)，免费与官方团队展开不同层次的合作，包括但不限于问题抽象、确定技术方案、项目答疑、共同研发等。如果您已经使用PaddleOCR落地项目，也可以填写此问卷，与飞桨平台共同宣传推广，提升企业技术品宣。期待您的提交！
<a href="https://trackgit.com">
<img src="https://us-central1-trackgit-analytics.cloudfunctions.net/token/ping/l63cvzo0w09yxypc7ygl" alt="traffic" />
</a>
applications/README_en.md 0 → 100644
View file @ c89bc397
English| [简体中文](README.md)
# Application
PaddleOCR scene application covers general, manufacturing, finance, transportation industry of the main OCR vertical applications, on the basis of the general capabilities of PP-OCR, PP-Structure, in the form of notebook to show the use of scene data fine-tuning, model optimization methods, data augmentation and other content, for developers to quickly land OCR applications to provide demonstration and inspiration.
- [Tutorial](#1)
- [General](#11)
- [Manufacturing](#12)
- [Finance](#13)
- [Transportation](#14)
- [Model Download](#2)
<a name="1"></a>
## Tutorial
<a name="11"></a>
### General
| Case | Feature | Model Download | Tutorial |
| ---------------------------------------------- | ---------------- | -------------------- | --------------------------------------- |
| High-precision Chineses recognition model SVTR | New model | [Model Download](#2) | [中文](./高精度中文识别模型.md)/English |
| Chinese handwriting recognition | New font support | | |
<a name="12"></a>
### Manufacturing
| Case | Feature | Model Download | Tutorial | Example |
| ------------------------------ | ------------------------------------------------------------ | -------------------- | ------------------------------------------------------------ | ------------------------------------------------------------ |
| Digital tube | Digital tube data sythesis, recognition model fine-tuning | [Model Download](#2) | [中文](./光功率计数码管字符识别/光功率计数码管字符识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/7d5774a273f84efba5b9ce7fd3f86e9ef24b6473e046444db69fa3ca20ac0986" width = "200" height = "100" /> |
| LCD screen | Detection model distillation, serving deployment | [Model Download](#2) | [中文](./液晶屏读数识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/901ab741cb46441ebec510b37e63b9d8d1b7c95f63cc4e5e8757f35179ae6373" width = "200" height = "100" /> |
| Packaging production data | Dot matrix character synthesis, overexposure and overdark text recognition | [Model Download](#2) | [中文](./包装生产日期识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/d9e0533cc1df47ffa3bbe99de9e42639a3ebfa5bce834bafb1ca4574bf9db684" width = "200" height = "100" /> |
| PCB text recognition | Small size text detection and recognition | [Model Download](#2) | [中文](./PCB字符识别/PCB字符识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/95d8e95bf1ab476987f2519c0f8f0c60a0cdc2c444804ed6ab08f2f7ab054880" width = "200" height = "100" /> |
| Meter text recognition | High-resolution image detection fine-tuning | [Model Download](#2) | | |
| LCD character defect detection | Non-text character recognition | | | |
<a name="13"></a>
### Finance
| Case | Feature | Model Download | Tutorial | Example |
| ----------------------------------- | --------------------------------------------- | -------------------- | ----------------------------------- | ------------------------------------------------------------ |
| Form visual question and answer | Multimodal general form structured extraction | [Model Download](#2) | [中文](./多模态表单识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/a3b25766f3074d2facdf88d4a60fc76612f51992fd124cf5bd846b213130665b" width = "200" height = "200" /> |
| VAT invoice | coming soon | | | |
| Seal detection and recognition | End-to-end curved text recognition | | | |
| Universal card recognition | Universal structured extraction | | | |
| ID card recognition | Structured extraction, image shading | | | |
| Contract key information extraction | Dense text detection, NLP concatenation | | | |
<a name="14"></a>
### Transportation
| Case | Feature | Model Download | Tutorial | Example |
| ----------------------------------------------- | ------------------------------------------------------------ | -------------------- | ----------------------------------- | ------------------------------------------------------------ |
| License plate recognition | Multi-angle images, lightweight models, edge-side deployment | [Model Download](#2) | [中文](./轻量级车牌识别.md)/English | <img src="https://ai-studio-static-online.cdn.bcebos.com/76b6a0939c2c4cf49039b6563c4b28e241e11285d7464e799e81c58c0f7707a7" width = "200" height = "100" /> |
| Driver's license/driving license identification | coming soon | | | |
| Express text recognition | coming soon | | | |
<a name="2"></a>
## Model Download
- For international developers: We're building a way to download these trained models, and since the current tutorials are Chinese, if you are good at both Chinese and English, or willing to polish English documents, please let us know in [discussion](https://github.com/PaddlePaddle/PaddleOCR/discussions).
- For Chinese developer: If you want to download the trained application model in the above scenarios, scan the QR code below with your WeChat, follow the PaddlePaddle official account to fill in the questionnaire, and join the PaddleOCR official group to get the 20G OCR learning materials (including "Dive into OCR" e-book, course video, application models and other materials)
<div align="center">
<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e" width = "150" height = "150" />
</div>
If you are an enterprise developer and have not found a suitable solution in the above scenarios, you can fill in the [OCR Application Cooperation Survey Questionnaire](https://paddle.wjx.cn/vj/QwF7GKw.aspx) to carry out different levels of cooperation with the official team **for free**, including but not limited to problem abstraction, technical solution determination, project Q&A, joint research and development, etc. If you have already used paddleOCR in your project, you can also fill out this questionnaire to jointly promote with the PaddlePaddle and enhance the technical publicity of enterprises. Looking forward to your submission!
<a href="https://trackgit.com">
<img src="https://us-central1-trackgit-analytics.cloudfunctions.net/token/ping/l6u6aszdfexs2jnrlil6" alt="trackgit-views" />
</a>
applications/光功率计数码管字符识别/corpus/digital.txt 0 → 100644
View file @ c89bc397
46.39
40.08
89.52
-71.93
23.19
-81.02
-34.09
05.87
-67.80
-51.56
-34.58
37.91
56.98
29.01
-90.13
35.55
66.07
-90.35
-50.93
42.42
21.40
-30.99
-71.78
25.60
-48.69
-72.28
-17.55
-99.93
-47.35
-64.89
-31.28
-90.01
05.17
30.91
30.56
-06.90
79.05
67.74
-32.31
94.22
28.75
51.03
-58.96
applications/光功率计数码管字符识别/光功率计数码管字符识别.md 0 → 100644
View file @ c89bc397
# 光功率计数码管字符识别
本案例将使用OCR技术自动识别光功率计显示屏文字，通过本章您可以掌握：
- PaddleOCR快速使用
- 数据合成方法
- 数据挖掘方法
- 基于现有数据微调
## 1. 背景介绍
光功率计（optical power meter ）是指用于测量绝对光功率或通过一段光纤的光功率相对损耗的仪器。在光纤系统中，测量光功率是最基本的，非常像电子学中的万用表；在光纤测量中，光功率计是重负荷常用表。
<img src="https://bkimg.cdn.bcebos.com/pic/a08b87d6277f9e2f999f5e3e1c30e924b899f35a?x-bce-process=image/watermark,image_d2F0ZXIvYmFpa2U5Mg==,g_7,xp_5,yp_5/format,f_auto" width="400">
目前光功率计缺少将数据直接输出的功能，需要人工读数。这一项工作单调重复，如果可以使用机器替代人工，将节约大量成本。针对上述问题，希望通过摄像头拍照->智能读数的方式高效地完成此任务。
为实现智能读数，通常会采取文本检测+文本识别的方案：
第一步，使用文本检测模型定位出光功率计中的数字部分；
第二步，使用文本识别模型获得准确的数字和单位信息。
本项目主要介绍如何完成第二步文本识别部分，包括：真实评估集的建立、训练数据的合成、基于 PP-OCRv3 和 SVTR_Tiny 两个模型进行训练，以及评估和推理。
本项目难点如下：
- 光功率计数码管字符数据较少，难以获取。
- 数码管中小数点占像素较少，容易漏识别。
针对以上问题， 本例选用 PP-OCRv3 和 SVTR_Tiny 两个高精度模型训练，同时提供了真实数据挖掘案例和数据合成案例。基于 PP-OCRv3 模型，在构建的真实评估集上精度从 52% 提升至 72%，SVTR_Tiny 模型精度可达到 78.9%。
aistudio项目链接: [光功率计数码管字符识别](https://aistudio.baidu.com/aistudio/projectdetail/4049044?contributionType=1)
## 2. PaddleOCR 快速使用
PaddleOCR 旨在打造一套丰富、领先、且实用的OCR工具库，助力开发者训练出更好的模型，并应用落地。
![](https://github.com/PaddlePaddle/PaddleOCR/raw/release/2.5/doc/imgs_results/ch_ppocr_mobile_v2.0/test_add_91.jpg)
官方提供了适用于通用场景的高精轻量模型，首先使用官方提供的 PP-OCRv3 模型预测图片，验证下当前模型在光功率计场景上的效果。
- 准备环境
```
python3 -m pip install -U pip
python3 -m pip install paddleocr
```
- 测试效果
测试图：
![](https://ai-studio-static-online.cdn.bcebos.com/8dca91f016884e16ad9216d416da72ea08190f97d87b4be883f15079b7ebab9a)
```
paddleocr --lang=ch --det=Fase --image_dir=data
```
得到如下测试结果：
```
('.7000', 0.6885431408882141)
```
发现数字识别较准，然而对负号和小数点识别不准确。 由于PP-OCRv3的训练数据大多为通用场景数据，在特定的场景上效果可能不够好。因此需要基于场景数据进行微调。
下面就主要介绍如何在光功率计（数码管）场景上微调训练。
## 3. 开始训练
### 3.1 数据准备
特定的工业场景往往很难获取开源的真实数据集，光功率计也是如此。在实际工业场景中，可以通过摄像头采集的方法收集大量真实数据，本例中重点介绍数据合成方法和真实数据挖掘方法，如何利用有限的数据优化模型精度。
数据集分为两个部分：合成数据，真实数据， 其中合成数据由 text_renderer 工具批量生成得到， 真实数据通过爬虫等方式在百度图片中搜索并使用 PPOCRLabel 标注得到。
- 合成数据
本例中数据合成工具使用的是 [text_renderer](https://github.com/Sanster/text_renderer)， 该工具可以合成用于文本识别训练的文本行数据:
![](https://github.com/oh-my-ocr/text_renderer/raw/master/example_data/effect_layout_image/char_spacing_compact.jpg)
![](https://github.com/oh-my-ocr/text_renderer/raw/master/example_data/effect_layout_image/color_image.jpg)
```
export https_proxy=http://172.19.57.45:3128
git clone https://github.com/oh-my-ocr/text_renderer
```
```
import os
python3 setup.py develop
python3 -m pip install -r docker/requirements.txt
python3 main.py \
--config example_data/example.py \
--dataset img \
--num_processes 2 \
--log_period 10
```
给定字体和语料，就可以合成较为丰富样式的文本行数据。 光功率计识别场景，目标是正确识别数码管文本，因此需要收集部分数码管字体，训练语料，用于合成文本识别数据。
将收集好的语料存放在 example_data 路径下：
```
ln -s ./fonts/DS* text_renderer/example_data/font/
ln -s ./corpus/digital.txt text_renderer/example_data/text/
```
修改 text_renderer/example_data/font_list/font_list.txt ,选择需要的字体开始合成：
```
python3 main.py \
--config example_data/digital_example.py \
--dataset img \
--num_processes 2 \
--log_period 10
```
合成图片会被存在目录 text_renderer/example_data/digital/chn_data 下
查看合成的数据样例：
![](https://ai-studio-static-online.cdn.bcebos.com/7d5774a273f84efba5b9ce7fd3f86e9ef24b6473e046444db69fa3ca20ac0986)
- 真实数据挖掘
模型训练需要使用真实数据作为评价指标，否则很容易过拟合到简单的合成数据中。没有开源数据的情况下，可以利用部分无标注数据+标注工具获得真实数据。
1. 数据搜集
使用[爬虫工具](https://github.com/Joeclinton1/google-images-download.git)获得无标注数据
2. [PPOCRLabel](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.5/PPOCRLabel) 完成半自动标注
PPOCRLabel是一款适用于OCR领域的半自动化图形标注工具，内置PP-OCR模型对数据自动标注和重新识别。使用Python3和PyQT5编写，支持矩形框标注、表格标注、不规则文本标注、关键信息标注模式，导出格式可直接用于PaddleOCR检测和识别模型的训练。
![](https://github.com/PaddlePaddle/PaddleOCR/raw/release/2.5/PPOCRLabel/data/gif/steps_en.gif)
收集完数据后就可以进行分配了，验证集中一般都是真实数据，训练集中包含合成数据+真实数据。本例中标注了155张图片，其中训练集和验证集的数目为100和55。
最终 `data` 文件夹应包含以下几部分：
```
|-data
|- synth_train.txt
|- real_train.txt
|- real_eval.txt
|- synthetic_data
|- word_001.png
|- word_002.jpg
|- word_003.jpg
| ...
|- real_data
|- word_001.png
|- word_002.jpg
|- word_003.jpg
| ...
...
```
### 3.2 模型选择
本案例提供了2种文本识别模型：PP-OCRv3 识别模型 和 SVTR_Tiny：
[PP-OCRv3 识别模型](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/PP-OCRv3_introduction.md)：PP-OCRv3的识别模块是基于文本识别算法SVTR优化。SVTR不再采用RNN结构，通过引入Transformers结构更加有效地挖掘文本行图像的上下文信息，从而提升文本识别能力。并进行了一系列结构改进加速模型预测。
[SVTR_Tiny](https://arxiv.org/abs/2205.00159):SVTR提出了一种用于场景文本识别的单视觉模型，该模型在patch-wise image tokenization框架内，完全摒弃了序列建模，在精度具有竞争力的前提下，模型参数量更少，速度更快。
以上两个策略在自建中文数据集上的精度和速度对比如下：
| ID | 策略 | 模型大小 | 精度 | 预测耗时（CPU + MKLDNN)|
|-----|-----|--------|----| --- |
| 01 | PP-OCRv2 | 8M | 74.8% | 8.54ms |
| 02 | SVTR_Tiny | 21M | 80.1% | 97ms |
| 03 | SVTR_LCNet(h32) | 12M | 71.9% | 6.6ms |
| 04 | SVTR_LCNet(h48) | 12M | 73.98% | 7.6ms |
| 05 | + GTC | 12M | 75.8% | 7.6ms |
| 06 | + TextConAug | 12M | 76.3% | 7.6ms |
| 07 | + TextRotNet | 12M | 76.9% | 7.6ms |
| 08 | + UDML | 12M | 78.4% | 7.6ms |
| 09 | + UIM | 12M | 79.4% | 7.6ms |
### 3.3 开始训练
首先下载 PaddleOCR 代码库
```
git clone -b release/2.5 https://github.com/PaddlePaddle/PaddleOCR.git
```
PaddleOCR提供了训练脚本、评估脚本和预测脚本，本节将以 PP-OCRv3 中文识别模型为例：
**Step1：下载预训练模型**
首先下载 pretrain model，您可以下载训练好的模型在自定义数据上进行finetune
```
cd PaddleOCR/
# 下载PP-OCRv3 中文预训练模型
wget -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar
# 解压模型参数
cd pretrain_models
tar -xf ch_PP-OCRv3_rec_train.tar && rm -rf ch_PP-OCRv3_rec_train.tar
```
**Step2：自定义字典文件**
接下来需要提供一个字典（{word_dict_name}.txt），使模型在训练时，可以将所有出现的字符映射为字典的索引。
因此字典需要包含所有希望被正确识别的字符，{word_dict_name}.txt需要写成如下格式，并以 `utf-8` 编码格式保存：
```
0
1
2
3
4
5
6
7
8
9
-
.
```
word_dict.txt 每行有一个单字，将字符与数字索引映射在一起，“3.14” 将被映射成 [3, 11, 1, 4]
* 内置字典
PaddleOCR内置了一部分字典，可以按需使用。
`ppocr/utils/ppocr_keys_v1.txt` 是一个包含6623个字符的中文字典
`ppocr/utils/ic15_dict.txt` 是一个包含36个字符的英文字典
* 自定义字典
内置字典面向通用场景，具体的工业场景中，可能需要识别特殊字符，或者只需识别某几个字符，此时自定义字典会更提升模型精度。例如在光功率计场景中，需要识别数字和单位。
遍历真实数据标签中的字符，制作字典`digital_dict.txt`如下所示：
```
-
.
0
1
2
3
4
5
6
7
8
9
B
E
F
H
L
N
T
W
d
k
m
n
o
z
```
**Step3：修改配置文件**
为了更好的使用预训练模型，训练推荐使用[ch_PP-OCRv3_rec_distillation.yml](../../configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml)配置文件，并参考下列说明修改配置文件：
以 `ch_PP-OCRv3_rec_distillation.yml` 为例：
```
Global:
...
# 添加自定义字典，如修改字典请将路径指向新字典
character_dict_path: ppocr/utils/dict/digital_dict.txt
...
# 识别空格
use_space_char: True
Optimizer:
...
# 添加学习率衰减策略
lr:
name: Cosine
learning_rate: 0.001
...
...
Train:
dataset:
# 数据集格式，支持LMDBDataSet以及SimpleDataSet
name: SimpleDataSet
# 数据集路径
data_dir: ./data/
# 训练集标签文件
label_file_list:
- ./train_data/digital_img/digital_train.txt #11w
- ./train_data/digital_img/real_train.txt #100
- ./train_data/digital_img/dbm_img/dbm.txt #3w
ratio_list:
- 0.3
- 1.0
- 1.0
transforms:
...
- RecResizeImg:
# 修改 image_shape 以适应长文本
image_shape: [3, 48, 320]
...
loader:
...
# 单卡训练的batch_size
batch_size_per_card: 256
...
Eval:
dataset:
# 数据集格式，支持LMDBDataSet以及SimpleDataSet
name: SimpleDataSet
# 数据集路径
data_dir: ./data
# 验证集标签文件
label_file_list:
- ./train_data/digital_img/real_val.txt
transforms:
...
- RecResizeImg:
# 修改 image_shape 以适应长文本
image_shape: [3, 48, 320]
...
loader:
# 单卡验证的batch_size
batch_size_per_card: 256
...
```
**注意，训练/预测/评估时的配置文件请务必与训练一致。**
**Step4：启动训练**
*如果您安装的是cpu版本，请将配置文件中的 `use_gpu` 字段修改为false*
```
# GPU训练 支持单卡，多卡训练
# 训练数码管数据 训练日志会自动保存为 "{save_model_dir}" 下的train.log
#单卡训练（训练周期长，不建议）
python3 tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model=./pretrain_models/ch_PP-OCRv3_rec_train/best_accuracy
#多卡训练，通过--gpus参数指定卡号
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model=./pretrain_models/en_PP-OCRv3_rec_train/best_accuracy
```
PaddleOCR支持训练和评估交替进行, 可以在 `configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml` 中修改 `eval_batch_step` 设置评估频率，默认每500个iter评估一次。评估过程中默认将最佳acc模型，保存为 `output/ch_PP-OCRv3_rec_distill/best_accuracy` 。
如果验证集很大，测试将会比较耗时，建议减少评估次数，或训练完再进行评估。
### SVTR_Tiny 训练
SVTR_Tiny 训练步骤与上面一致，SVTR支持的配置和模型训练权重可以参考[算法介绍文档](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.5/doc/doc_ch/algorithm_rec_svtr.md)
**Step1：下载预训练模型**
```
# 下载 SVTR_Tiny 中文识别预训练模型和配置文件
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/rec_svtr_tiny_none_ctc_ch_train.tar
# 解压模型参数
tar -xf rec_svtr_tiny_none_ctc_ch_train.tar && rm -rf rec_svtr_tiny_none_ctc_ch_train.tar
```
**Step2：自定义字典文件**
字典依然使用自定义的 digital_dict.txt
**Step3：修改配置文件**
配置文件中对应修改字典路径和数据路径
**Step4：启动训练**
```
## 单卡训练
python tools/train.py -c rec_svtr_tiny_none_ctc_ch_train/rec_svtr_tiny_6local_6global_stn_ch.yml \
-o Global.pretrained_model=./rec_svtr_tiny_none_ctc_ch_train/best_accuracy
```
### 3.4 验证效果
如需获取已训练模型，请扫码填写问卷，加入PaddleOCR官方交流群获取全部OCR垂类模型下载链接、《动手学OCR》电子书等全套OCR学习资料🎁
<div align="left">
<img src="https://ai-studio-static-online.cdn.bcebos.com/dd721099bd50478f9d5fb13d8dd00fad69c22d6848244fd3a1d3980d7fefc63e" width = "150" height = "150" />
</div>
将下载或训练完成的模型放置在对应目录下即可完成模型推理
* 指标评估
训练中模型参数默认保存在`Global.save_model_dir`目录下。在评估指标时，需要设置`Global.checkpoints`指向保存的参数文件。评估数据集可以通过 `configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml` 修改Eval中的 `label_file_path` 设置。
```
# GPU 评估， Global.checkpoints 为待测权重
python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.checkpoints={path/to/weights}/best_accuracy
```
* 测试识别效果
使用 PaddleOCR 训练好的模型，可以通过以下脚本进行快速预测。
默认预测图片存储在 `infer_img` 里，通过 `-o Global.checkpoints` 加载训练好的参数文件：
根据配置文件中设置的 `save_model_dir` 和 `save_epoch_step` 字段，会有以下几种参数被保存下来：
```
output/rec/
├── best_accuracy.pdopt
├── best_accuracy.pdparams
├── best_accuracy.states
├── config.yml
├── iter_epoch_3.pdopt
├── iter_epoch_3.pdparams
├── iter_epoch_3.states
├── latest.pdopt
├── latest.pdparams
├── latest.states
└── train.log
```
其中 best_accuracy.* 是评估集上的最优模型；iter_epoch_x.* 是以 `save_epoch_step` 为间隔保存下来的模型；latest.* 是最后一个epoch的模型。
```
# 预测英文结果
python3 tools/infer_rec.py -c configs/rec/PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.infer_img=test_digital.png
```
预测图片：
![](https://ai-studio-static-online.cdn.bcebos.com/8dca91f016884e16ad9216d416da72ea08190f97d87b4be883f15079b7ebab9a)
得到输入图像的预测结果：
```
infer_img: test_digital.png
result: ('-70.00', 0.9998967)
```
applications/包装生产日期识别.md 0 → 100644
View file @ c89bc397
This diff is collapsed.
applications/多模态表单识别.md
View file @ c89bc397
This diff is collapsed.
applications/液晶屏读数识别.md 0 → 100644
View file @ c89bc397
This diff is collapsed.
applications/轻量级车牌识别.md 0 → 100644
View file @ c89bc397
This diff is collapsed.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment