JessicaSZX / ocr-digital-display

This optical character recognition program is for identifying contents on seven segment display and bitmap.

This deep learning OCR project is aiming to recognize digital display contents which includes seven segment and bitmpap.

Please note that training is conducing under gluon-cv/docs/tutorials/detection/ from official gluon-cv repo: https://github.com/dmlc/gluon-cv

STEP1: Dataset preparation:

- Generate customized dataset
    cd ocr-ledsegment-bitmap/text_gen/TextRecognitionDataGenerator
    python run.py -f 64 -l num -c 8000 --length 8 -k 5 -wd -5 -tc '#000000','#999999' -m 20,20,20,20 -bl 0
    python gen_voc_ann.py

       
- Make sure here is no VOC2019 dataset already exist 
    cd ocr-ledsegment-bitmap/
    rm -r VOC2019

- Generate VOC dataset
    python make_voc.py
    python make_train_val_test.py

STEP2: Pass dataset to Network:

- Make sure here is no VOC2019 dataset already exist
    cd gluon-cv/docs/tutorials/detection/
    rm -r VOC2019
    cp -r ~/ocr-ledsegment-bitmap/VOC2019/ ./

STEP3: Training:

    source /opt/intel/compilers_and_libraries/linux/bin/compilervars.sh intel64
    export PYTHONPATH=~/MXNet-MKL-DNN/incubator-mxnet/python/

    rm -r number_train_result/
    python train_ssd.py --save-prefix number_train_result2/ --epochs 70 --batch-size 32 --lr 0.001 --lr-decay 0.1 --lr-decay-epoch 40,50,60