text_recognition_toolbox: The reimplementation of a series of classical scene text recognition papers with Pytorch in a uniform way.

Overview

text recognition toolbox

1. 项目介绍

该项目是基于pytorch深度学习框架,以统一的改写方式实现了以下6篇经典的文字识别论文,论文的详情如下。该项目会持续进行更新,欢迎大家提出问题以及对代码进行贡献。

模型 论文标题 发表年份 模型方法划分
CRNN 《An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition》 2017 CNN+BiLSTM+CTC
GRCNN 《Gated recurrent convolution neural network for OCR》 2017 Gated Recurrent Convulution Layer + BiSTM + CTC
FAN 《Focusing attention: Towards accurate text recognition in natural images》 2017 focusing network+1D attention
SAR 《Show, attend and read: A simple and strong baseline for irregular text recognition》 2019 ResNet+2D attention
DAN 《Decoupled attention network for text recognition》 2020 FCN+convolutional alignment module
SATRN 《On Recognizing Texts of Arbitrary Shapes with 2D Self-Attention》 2020 Transformer

2. 如何使用

2.1 环境要求

torch==1.3.0
numpy==1.17.3
lmdb==0.98
opencv-python==3.4.5.20

2.2 训练

  • 数据准备

首先需要准备训练数据,目前只支持lmdb格式的数据,数据转换的步骤如下:

  1. 准备图片数据集,图片是根据检测框进行切分后的数据
  2. 准备label.txt,标注文件需保持如下的格式
1.jpg 文字检测
2.jpg 文字识别
  1. 进行lmdb格式数据集的转换
python3 tools/create_lmdb_dataset.py --inputPath {图片数据集路径} --gtFile {标注文件路径} --outputPath {lmdb格式数据集保存路径}
  • 配置文件

目前每个模型都单独配备了一个配置文件,这里以CRNN为例, 配置文件主要参数的含义如下:

一级参数 二级参数 参数含义 备注
TrainReader dataloader 自定义的DataLoader类
select_data 选择使用的lmdb格式数据集 默认为'/',即使用{lmdb_sets_dir}路径下所有的lmdb数据集。如果想控制同一个batch里不同数据集的比例,可以配合{batch_ratio}使用,并将数据集名称用'-'进行分割,例如设置成'数据集1-数据集2-数据集3'
batch_ratio 控制在一个batch中,各个lmdb格式数据集的比例 配合{select_data}进行使用,将比例用'-'进行分割,例如设置成'0.3-0.3-0.4'。即数据集1使用batch_size * 0.3的比例,剩余的数据集以此类推。
total_data_usage_ratio 控制使用的整体数据集比例 默认为1.0,即使用全部的数据集
padding 是否对数据进行padding补齐 默认为True,设置为False即采用resize的方式
Global highest_acc_save_type 是否只保存识别率最高的模型 默认为False
resumed_optimizer 是否加载之前保存的optimizer 默认为False
batch_max_length 最大的字符串长度 超过这个字符串长度的训练数据会被过滤掉
eval_batch_step 保存模型的间隔步数
Architecture function 使用的模型 此处为'CRNN'
SeqRNN input_size LSTM输入的尺寸 即backbone输出的通道个数
hidden_size LSTM隐藏层的尺寸
  • 模型训练

完成上述配置后,使用以下命令即可开始模型的训练:

python train.py -c configs/CRNN.yml

2.3 预测

  • 配置文件

同样地,针对模型预测,也都单独配备了一个配置文件,这里以CRNN为例, 需要修改的配置参数如下:

一级参数 二级参数 参数含义 备注
Global pretrain_weights 模型文件路径 剩余配置参数和训练保持一致即可
infer_img 待预测的图片,可以是文件夹或者是图片路径
  • 模型预测

完成上述配置后,使用以下命令即可开始模型的预测:

python predict.py -c configs/CRNN.yml

3. 预训练模型

以下是5个开源的中文自然场景数据集,可以直接根据上述的模型配置进行模型训练:

数据集 网盘地址 备注
一共包括5个自然场景训练集:
ArT_train, LSVT_train, MTWI_train, RCTW17_train, ReCTS_train
以及一个自然场景验证集:ReCTS_val
链接: https://pan.baidu.com/s/1fvExHzeojA_Yhj3_wDflwA
提取码: kzrd
"train"是训练集,"val"是验证集

以下为5个算法的预训练模型,训练的明细请见第4部分里的实验设定:

模型 网盘地址 备注
一共包含5个预训练模型:CRNN.pth, GRCNN.pth, FAN.pth, DAN.pth, SAR.pth
以及一个字典文件:keys.txt
链接: https://pan.baidu.com/s/1IG-1lxytrOqry9c5Nc1GzQ
提取码: k3ij

4. 实验结果

针对目前已复现的5个算法,我用统一的数据集以及参数设定进行了实验对比,实验设定以及实验结果如下:

  • 实验设定
实验设定 明细 备注
训练集 ArT_train:44663
LSVT_train:218552
MTWI_train:79964
RCTW17_train:33342
ReCTS_train:83119
这5个均为开源自然场景数据集,其中做了剔除模糊数据等处理
验证集 ReCTS_val:9231 测试集为从ReCTS中按照9:1比例划分的验证集,注意ReCTS以水平文本居多
batch_size 128
img_shape [1, 32, 256] 尺寸进行等比例放缩,小于256的进行padding,大于256的resize至256
optimizer function: adam
base_lr: 0.001
momentum: 0.9
weight_decay: 1.0e-4
iter 60000 一共训练了60000步,每2000步会进行一次验证
  • 实验结果
算法 最高识别率 最大正则编辑距离 模型大小
CRNN 59.89 0.7959 120M
GRCNN 70.51 0.8597 78M
FAN 75.78 0.8924 764M
SAR 78.13 0.9037 722M
DAN 78.99 0.9064 639M

下图为各个算法在验证集上的识别率,每2000步会进行验证:

fig1

  • 预测结果示例
算法 预测结果 备注
CRNN image-20210121152011971 预测结果均取自验证集识别率最高的模型,
左边一列为预测结果,右边为标注结果
GRCNN image-20210121152134249
FAN image-20210121152239497
SAR image-20210121152325124
DAN image-20210121152407344
CS550 Machine Learning course project on CNN Detection.

CNN Detection (CS550 Machine Learning Project) Team Members (Tensor) : Yadava Kishore Chodipilli (11940310) Thashmitha BS (11941250) This is a work do

yaadava_kishore 2 Jan 30, 2022
SelfRemaster: SSL Speech Restoration

SelfRemaster: Self-Supervised Speech Restoration Official implementation of SelfRemaster: Self-Supervised Speech Restoration with Analysis-by-Synthesi

Takaaki Saeki 46 Jan 07, 2023
Official DGL implementation of "Rethinking High-order Graph Convolutional Networks"

SE Aggregation This is the implementation for Rethinking High-order Graph Convolutional Networks. Here we show the codes for citation networks as an e

Tianqi Zhang (张天启) 32 Jul 19, 2022
POT : Python Optimal Transport

POT: Python Optimal Transport This open source Python library provide several solvers for optimization problems related to Optimal Transport for signa

Python Optimal Transport 1.7k Dec 31, 2022
EGNN - Implementation of E(n)-Equivariant Graph Neural Networks, in Pytorch

EGNN - Pytorch Implementation of E(n)-Equivariant Graph Neural Networks, in Pytorch. May be eventually used for Alphafold2 replication. This

Phil Wang 259 Jan 04, 2023
Companion code for the paper Theoretical characterization of uncertainty in high-dimensional linear classification

Companion code for the paper Theoretical characterization of uncertainty in high-dimensional linear classification Usage The required packages are lis

0 Feb 07, 2022
The authors' official PyTorch SigWGAN implementation

The authors' official PyTorch SigWGAN implementation This repository is the official implementation of [Sig-Wasserstein GANs for Time Series Generatio

9 Jun 16, 2022
Apache Flink

Apache Flink Apache Flink is an open source stream processing framework with powerful stream- and batch-processing capabilities. Learn more about Flin

The Apache Software Foundation 20.4k Dec 30, 2022
LOFO (Leave One Feature Out) Importance calculates the importances of a set of features based on a metric of choice,

LOFO (Leave One Feature Out) Importance calculates the importances of a set of features based on a metric of choice, for a model of choice, by iteratively removing each feature from the set, and eval

Ahmet Erdem 691 Dec 23, 2022
Chainer Implementation of Semantic Segmentation using Adversarial Networks

Semantic Segmentation using Adversarial Networks Requirements Chainer (1.23.0) Differences Use of FCN-VGG16 instead of Dilated8 as Segmentor. Caution

Taiki Oyama 99 Jun 28, 2022
Custom studies about block sparse attention.

Block Sparse Attention 研究总结 本人近半年来对Block Sparse Attention(块稀疏注意力)的研究总结(持续更新中)。按时间顺序,主要分为如下三部分: PyTorch 自定义 CUDA 算子——以矩阵乘法为例 基于 Triton 的 Block Sparse A

Chen Kai 2 Jan 09, 2022
PlenOctree Extraction algorithm

PlenOctrees_NeRF-SH This is an implementation of the Paper PlenOctrees for Real-time Rendering of Neural Radiance Fields. Not only the code provides t

49 Nov 05, 2022
Retrieval.pytorch - The code we used in [2020 DIGIX]

Retrieval.pytorch - The code we used in [2020 DIGIX]

Guo-Hua Wang 2 Feb 07, 2022
Efficient Training of Visual Transformers with Small Datasets

Official codes for "Efficient Training of Visual Transformers with Small Datasets", NerIPS 2021.

Yahui Liu 112 Dec 25, 2022
Facebook AI Research Sequence-to-Sequence Toolkit written in Python.

Fairseq(-py) is a sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language mod

20.5k Jan 08, 2023
Official implementation of the paper Image Generators with Conditionally-Independent Pixel Synthesis https://arxiv.org/abs/2011.13775

CIPS -- Official Pytorch Implementation of the paper Image Generators with Conditionally-Independent Pixel Synthesis Requirements pip install -r requi

Multimodal Lab @ Samsung AI Center Moscow 201 Dec 21, 2022
This package implements THOR: Transformer with Stochastic Experts.

THOR: Transformer with Stochastic Experts This PyTorch package implements Taming Sparsely Activated Transformer with Stochastic Experts. Installation

Microsoft 45 Nov 22, 2022
Point-NeRF: Point-based Neural Radiance Fields

Point-NeRF: Point-based Neural Radiance Fields Project Sites | Paper | Primary c

Qiangeng Xu 662 Jan 01, 2023
Official Code for "Non-deep Networks"

Non-deep Networks arXiv:2110.07641 Ankit Goyal, Alexey Bochkovskiy, Jia Deng, Vladlen Koltun Overview: Depth is the hallmark of DNNs. But more depth m

Ankit Goyal 567 Dec 12, 2022
A Jinja extension (compatible with Flask and other frameworks) to compile and/or compress your assets.

A Jinja extension (compatible with Flask and other frameworks) to compile and/or compress your assets.

Jayson Reis 94 Nov 21, 2022