Chinese Character Structure Detection

🏆 The third prize of the 14th National College Student Software Innovation Competition.

Overall Procedure

Model

Prepare Dataset

1. Put the the annotated dataset in data directory.
2. run python -m src.preprocess.peanut

We have annotated part of the Chinese handwriting character dataset, please refer to ccsa for more details.

Train

Run base network

python -m src.preprocess.run 'train' 'base' --batch-size=16 --num-workers=0 --epoch=50 --step-size=5 --lr=3e-4 --use-gpu --gpu=0 --apex

Run refinement network

python -m src.preprocess.run 'train' 'refinement' --batch-size=16 --num-workers=0 --epoch=10 --step-size=5 --lr=5e-3 --use-gpu --gpu=0 --apex

Run classification network to get a pretrained model

python -m src.preprocess.run 'train' 'cls' --batch-size=16 --num-workers=0 --epoch=10 --step-size=5 --lr=5e-3 --use-gpu --gpu=0 --apex

Inference

Run inference model, using --snapshot to specify model weights

python -m src.preprocess.run 'test' 'openpose' --batch-size=32 --num-workers=0 --use-gpu --gpu=2 --apex --reinforcement --snapshot='./checkpoints/openpose_model_apex_ep2.pt'

Visualize Results

python -m src.preprocess.run 'view' 'base'

Results

Heatmap Results

Character Structure Detection

Contributors

Team Members: Chuyang, QER, LBY, XRZ