GitPlanet

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

Created with love in Canada, visit hostnodejs.com today

Feel like to post an Ad? Learn Details
All Projects → BADBADBADBOY → Dbnet Lite.pytorch

BADBADBADBOY / Dbnet Lite.pytorch

A pytorch re-implementation of Real-time Scene Text Detection with Differentiable Binarization

Programming Languages

python
139335 projects - #7 most used programming language
python3
1442 projects

Labels

pytorchocrlightweight

Projects that are alternatives of or similar to Dbnet Lite.pytorch

Simplehtr
Handwritten Text Recognition (HTR) system implemented with TensorFlow.
Stars: ✭ 1,072 (+1453.62%)
Mutual labels:  ocr
Infrared
An ultra lightweight minecraft reverse proxy and idle placeholder
Stars: ✭ 59 (-14.49%)
Mutual labels:  lightweight
Esa Restlight
ESA Restlight is a lightweight and rest-oriented web framework.
Stars: ✭ 67 (-2.9%)
Mutual labels:  lightweight
Simiki
Simiki is a simple wiki framework, written in Python.
Stars: ✭ 1,082 (+1468.12%)
Mutual labels:  lightweight
Qtwebserver
Qt based web application server
Stars: ✭ 56 (-18.84%)
Mutual labels:  lightweight
Cnn lstm ctc ocr for icpr
Forked from weinman/cnn_lstm_ctc_ocr for the ICPR MTWI 2018 challenge 1
Stars: ✭ 62 (-10.14%)
Mutual labels:  ocr
Idmatch
Match faces on id cards with OCR capabilities.
Stars: ✭ 52 (-24.64%)
Mutual labels:  ocr
Face Recognition Cpp
Real Time Face Recognition Detector. Over 30FPS on CPU!
Stars: ✭ 68 (-1.45%)
Mutual labels:  lightweight
Sane Scan Pdf
Sane command-line scan-to-pdf script on Linux with OCR and deskew support
Stars: ✭ 58 (-15.94%)
Mutual labels:  ocr
Ultimatemrz Sdk
Machine-readable zone/travel document (MRZ / MRTD) detector and recognizer using deep learning
Stars: ✭ 66 (-4.35%)
Mutual labels:  ocr
Miniplayer
A curses based mpd client with basic functionality and album art.
Stars: ✭ 56 (-18.84%)
Mutual labels:  lightweight
Termly.js
Simple, Extensible, Hackable and Lightweight Javascript Browser Terminal Simulator!
Stars: ✭ 56 (-18.84%)
Mutual labels:  lightweight
Skautoscrolllabel
Automatically scrolling UILabel with both horizontal/vertical MARQUEE effects and gradient gradients on the edges. Gradient fading is used on the edge of the scroll to solve the problem of the hard edges of the rolling edge. The overall effect is a natural and easy to use.
Stars: ✭ 64 (-7.25%)
Mutual labels:  lightweight
Open Semantic Search Apps
Python/Django based webapps and web user interfaces for search, structure (meta data management like thesaurus, ontologies, annotations and named entities) and data import (ETL like text extraction, OCR and crawling filesystems or websites)
Stars: ✭ 55 (-20.29%)
Mutual labels:  ocr
Ocr Electron Vue
📇 A Simple OCR Application built on Electron, Vue.js & Tesseract.js
Stars: ✭ 67 (-2.9%)
Mutual labels:  ocr
Seven Segment Ocr
Digitizes numbers from videos & images of seven segment displays
Stars: ✭ 53 (-23.19%)
Mutual labels:  ocr
Doccreator
DIAR software for synthetic document image and groundtruth generation, with various degradation models for data augmentation
Stars: ✭ 60 (-13.04%)
Mutual labels:  ocr
Crnn
Convolutional recurrent neural network for scene text recognition or OCR in Keras
Stars: ✭ 68 (-1.45%)
Mutual labels:  ocr
Scanbot Sdk Example Android
Document scanning SDK example apps for the Scanbot SDK for Android.
Stars: ✭ 67 (-2.9%)
Mutual labels:  ocr
Ttkocr
OCR 识别图片以及PDF中的文字， based on Qt for windows and linux.
Stars: ✭ 65 (-5.8%)
Mutual labels:  ocr
View All Similar Projects ➔

English | 简体中文

DBNet-lite-pytorch

这个项目之后会在这里更新,我把之前的项目都做了下整合pytorchOCR

环境配置

pip install -r requirement.txt
cd models/dcn/
sh make.sh

水平或倾斜文本格式

照着icdar2015的格式, x1,y1,x2,y2,x3,y3,x4,y4,label,


image
│   1.jpg
│   2.jpg   
│		...
label
│   gt_1.txt
│   gt_2.txt
|		...

弧形文本的格式

数据格式, x1,y1,x2,y2,x3,y3,x4,y4 ...xn,yn,label

n个点组成，n的个数可以不定

image
│   1.jpg
│   2.jpg   
│		...
label
│   gt_1.txt
│   gt_2.txt
|		...

训练部分

在根目录的config.yaml里配置train部分的一些参数，例如一些图片位置，如果你的图片和gt文件名字是一样的，你可以设置is_icdar2015=False。 如果你不想做验证，可以直接设置start_val_epoch大于n_epoch，如果设置了做验证，会保存一个hmean最高的最优模型。

python3 train.py

测试部分

测试时，配置config.yaml中test部分，对于弧形文本设置is_poly=True,其它非弧形文本设置为False

python3 inference.py

模型压缩之通道剪裁

训练部分

  1. 先进行稀疏训练，首先修改config.yaml将use_sr 设置为True，并设定sr_lr，这个设置越大压的越多，注意设置太大有可能不收敛.
python3 train.py
  1. 压缩模型 设置好config.yaml中pruned部分参数，运行
python3 ./pruned/prune.py
  1. 重新finetune模型 这里精度会很快回升，一般可以训练50-100epoch，具体自己做实验
python3 ./pruned/train_fintune.py

测试部分

python3 ./pruned/prune_inference.py

在icdar2015的测试结果

Method head extra data prune ratio model size(M) precision(%) recall(%) hmean(%) model_file
Resnet18 FPN no 0 62.6 86.11 76.45 80.99 baiduyun (extract code: p0bk)
Resnet18 DB no 0.8 20.1 85.55 76.40 80.72

在icdar2015的测试结果图

该项目会做

  • [x] 转换作者的代码便于阅读和调试
  • [x] 展示一些训练结果
  • [ ] 加入轻量化的backbone压缩模型
  • [ ] 通过通道剪裁压缩DB模型，精度基本不变
  • [ ] 通过知识蒸馏进一步提升压缩后模型效果

参考

  1. https://github.com/whai362/PSENet
  2. https://github.com/MhLiao/DB
  3. https://github.com/Jzz24/pytorch_quantization
Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].