DocTrPP: DocTr++ in PaddlePaddle

Introduction

This is a PaddlePaddle implementation of DocTr++. The original paper is DocTr++: Deep Unrestricted Document Image Rectification. The original code is here.

Requirements

You need to install the latest version of PaddlePaddle, which is done through this link.

Training

1. Data Preparation

To prepare datasets, refer to doc3D.

2. Training

sh train.sh

or

export OPENCV_IO_ENABLE_OPENEXR=1
export CUDA_VISIBLE_DEVICES=0

python train.py --img-size 288 \
    --name "DocTr++" \
    --batch-size 12 \
    --lr 2.5e-5 \
    --exist-ok \
    --use-vdl

3. Load Trained Model and Continue Training

export OPENCV_IO_ENABLE_OPENEXR=1
export CUDA_VISIBLE_DEVICES=0

python train.py --img-size 288 \
    --name "DocTr++" \
    --batch-size 12 \
    --lr 2.5e-5 \
    --resume "runs/train/DocTr++/weights/last.ckpt" \
    --exist-ok \
    --use-vdl

Test and Inference

Test the dewarp result on a single image:

python predict.py -i "crop/12_2 copy.png" -m runs/train/DocTr++/weights/best.ckpt -o 12.2.png

Export to onnx

pip install paddle2onnx

python export.py -m ./best.ckpt --format onnx

Model Download

The trained model can be downloaded from here.