Skip to content

Latest commit

 

History

History
139 lines (95 loc) · 5.65 KB

algorithm_rec_starnet.md

File metadata and controls

139 lines (95 loc) · 5.65 KB

STAR-Net

1. 算法简介

论文信息:

STAR-Net: a spatial attention residue network for scene text recognition.

Wei Liu, Chaofeng Chen, Kwan-Yee K. Wong, Zhizhong Su and Junyu Han.

BMVC, pages 43.1-43.13, 2016

参考DTRB 文字识别训练和评估流程,使用MJSynth和SynthText两个文字识别数据集训练,在IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE数据集上进行评估,算法效果如下:

模型 骨干网络 Avg Accuracy 配置文件 下载链接
StarNet Resnet34_vd 84.44% configs/rec/rec_r34_vd_tps_bilstm_ctc.yml 训练模型
StarNet MobileNetV3 81.42% configs/rec/rec_mv3_tps_bilstm_ctc.yml 训练模型

2. 环境配置

请先参考《运行环境准备》配置PaddleOCR运行环境,参考《项目克隆》克隆项目代码。

3. 模型训练、评估、预测

请参考文本识别训练教程。PaddleOCR对代码进行了模块化,训练不同的识别模型只需要更换配置文件即可。

  • 训练

在完成数据准备后,便可以启动训练,训练命令如下:

#单卡训练(训练周期长,不建议)
python3 tools/train.py -c configs/rec/rec_r34_vd_tps_bilstm_ctc.yml

#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 rec_r34_vd_tps_bilstm_ctc.yml

  • 评估
# GPU 评估, Global.pretrained_model 为待测权重
python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/rec_r34_vd_tps_bilstm_ctc.yml -o Global.pretrained_model={path/to/weights}/best_accuracy
  • 预测:
# 预测使用的配置文件必须与训练一致
python3 tools/infer_rec.py -c configs/rec/rec_r34_vd_tps_bilstm_ctc.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.infer_img=doc/imgs_words/en/word_1.png

4. 推理部署

4.1 Python推理

首先将 STAR-Net 文本识别训练过程中保存的模型,转换成inference model。以基于Resnet34_vd骨干网络,使用MJSynth和SynthText两个英文文本识别合成数据集训练的模型 为例,可以使用如下命令进行转换:

python3 tools/export_model.py -c configs/rec/rec_r34_vd_tps_bilstm_ctc.yml -o Global.pretrained_model=./rec_r34_vd_tps_bilstm_ctc_v2.0_train/best_accuracy  Global.save_inference_dir=./inference/rec_starnet

STAR-Net 文本识别模型推理,可以执行如下命令:

python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words_en/word_336.png" --rec_model_dir="./inference/rec_starnet/" --rec_image_shape="3, 32, 100" --rec_char_dict_path="./ppocr/utils/ic15_dict.txt"

执行命令后,上面图像的识别结果如下:

Predicts of ./doc/imgs_words_en/word_336.png:('super', 0.9999073)

注意:由于上述模型是参考DTRB文本识别训练和评估流程,与超轻量级中文识别模型训练有两方面不同:

  • 训练时采用的图像分辨率不同,训练上述模型采用的图像分辨率是[3,32,100],而中文模型训练时,为了保证长文本的识别效果,训练时采用的图像分辨率是[3, 32, 320]。预测推理程序默认的形状参数是训练中文采用的图像分辨率,即[3, 32, 320]。因此,这里推理上述英文模型时,需要通过参数rec_image_shape设置识别图像的形状。

  • 字符列表,DTRB论文中实验只是针对26个小写英文本母和10个数字进行实验,总共36个字符。所有大小字符都转成了小写字符,不在上面列表的字符都忽略,认为是空格。因此这里没有输入字符字典,而是通过如下命令生成字典.因此在推理时需要设置参数rec_char_dict_path,指定为英文字典"./ppocr/utils/ic15_dict.txt"。

self.character_str = "0123456789abcdefghijklmnopqrstuvwxyz"
dict_character = list(self.character_str)

4.2 C++推理

准备好推理模型后,参考cpp infer教程进行操作即可。

4.3 Serving服务化部署

准备好推理模型后,参考pdserving教程进行Serving服务化部署,包括Python Serving和C++ Serving两种模式。

4.4 更多推理部署

STAR-Net模型还支持以下推理部署方式:

  • Paddle2ONNX推理:准备好推理模型后,参考paddle2onnx教程操作。

5. FAQ

引用

@inproceedings{liu2016star,
  title={STAR-Net: a spatial attention residue network for scene text recognition.},
  author={Liu, Wei and Chen, Chaofeng and Wong, Kwan-Yee K and Su, Zhizhong and Han, Junyu},
  booktitle={BMVC},
  volume={2},
  pages={7},
  year={2016}
}