Skip to content

Latest commit

 

History

History
80 lines (66 loc) · 3.45 KB

recognition.md

File metadata and controls

80 lines (66 loc) · 3.45 KB
Raw

TEXT RECOGNITION

This section uses the icdar2015 dataset as an example to introduce the training, evaluation, and testing of the recognition model in Gluon-OCR.

Data preparation

Downloading icdar2015 dataset from official website.

Decompress the downloaded dataset to the working directory, assuming it is decompressed under gluon-ocr/train_data/. After decompressing the data set and downloading the annotation file, gluon-ocr/train_data/ has two folders and two files, which are:

~/gluon-ocr/train_data/icdar2015/
  └─ ch4_train_word_imgs/                Training image of icdar dataset
  └─ ch4_test_word_imgs/                 Testing image of icdar dataset
  └─ train_icdar2015_label.txt           Training annotation of icdar dataset
  └─ test_icdar2015_label.txt/           Testing annotation of icdar dataset

The provided annotation file format is as follow, seperated by "\t", image file path is absolute path:

" Image file path                         Image annotation"
ch4_test_word_imgs/img_1.png                   JOINT
ch4_test_word_imgs/img_2.png                   yourself
        ...                                     ...

A dictionary ({word_dict_name}.txt) needs to be provided so that when the model is trained, all the characters that appear can be mapped to the dictionary index.

Therefore, the dictionary needs to contain all the characters that you want to be recognized correctly. {word_dict_name}.txt needs to be written in the following format and saved in the utf-8 encoding format:

l
d
a
...

~/gluon-ocr/gluonocr/utils/gluonocr_dict.txt is a Chinese dictionary with 7434 characters.

~/gluon-ocr/gluonocr/utils/ic15_dict.txt is an English dictionary with 36 characters.

Training

Training CRNN.

cd gluon-ocr/scripts/recog/crnn
python train_crnn.py --network resnet --num-layers 34 --batch-size 300 --dataset-name icdar15 \
--max-len 30 --fix-width 384 --short-side 32 --gpus 2,3 --num-workers 16 --warmup-epochs 0 \
--lr 0.001 --lr-decay 0.1 --lr-decay-epoch 60 --wd 0.0005 --voc-path ../../../gluonocr/utils/ic15_dict.txt \
--train-data-path ../../../train_data/icdar2015/train_icdar2015_label.txt \
--val-data-path ../../../train_data/icdar2015/test_icdar2015_label.txt \
--save-prefix ./checkpoint

Export model

cd gluon-ocr/scripts/recog/crnn
python train_crnn.py --network resnet --num-layers 34 --batch-size 300 --dataset-name icdar15 \
--max-len 30 --fix-width 384 --short-side 32 --gpus 2,3 --num-workers 16 --warmup-epochs 0 \
--lr 0.001 --lr-decay 0.1 --lr-decay-epoch 60 --wd 0.0005 --voc-path ../../../gluonocr/utils/ic15_dict.txt \
--train-data-path ../../../train_data/icdar2015/train_icdar2015_label.txt \
--val-data-path ../../../train_data/icdar2015/test_icdar2015_label.txt \
--save-prefix ./checkpoint --resume ./checkpoint/icdar15-resnet34-crnn_best.params --export-model

gluon-ocr/scripts/recog/crnn/checkpoint will be generated two files, which are:

  └─ icdar15-resnet34-crnn-0000.params          DBNet params file
  └─ icdar15-resnet34-crnn-symbol.json          DBNet model file

Test

Test the recognition result on a single image:

cd gluon-ocr/scripts/recog/crnn
python demo.py --model-path ./checkpoint/icdar15-resnet34-crnn-symbol.json \
--params-path ./checkpoint/icdar15-resnet34-crnn-0000.params \
--voc-path ../../../gluonocr/utils/ic15_dict.txt \
--image-path ../../../doc/imgs/word_1.png \
--gpu-id 1