Introduction

LDRNet is lightweight deep neural network for localizing document in a image/photo with extremely low latency. Given any image as input, LDRNet predicts the coordinates of the document's quadrilateral corners in real-time.

Requirements

• tensorflow==2.3.1
• tensorflow-addons==0.11.2
• opencv-python==4.4.0.46
• numpy==1.18.5

Predict

A simple example for predicting the quadrilateral corners of the image in imgs folder:

python3 predict.py

Train

CUDA_VSIBLE_DEVICES=0 python3 train.py --config_file ./config.yaml

• fill the label_path and img_folder_path in config.yaml
• label consists of 18 elements: file name, eights floats for four coordinates, eights float for weights(set to 0.0625), class index(start from 0)

Deployment

• To TensorRT
• To TNN

Citations

If you publish work related to LDRNet or Real-time Document Localization, please refer to the following paper and cite accordingly:

@inproceedings{wu2023ldrnet,
  title={LDRNet: Enabling Real-time Document Localization on Mobile Devices},
  author={Wu, Han and Qian, Holland and Wu, Huaming and van Moorsel, Aad},
  booktitle={Machine Learning and Principles and Practice of Knowledge Discovery in Databases: International Workshops of ECML PKDD 2022, Grenoble, France, September 19--23, 2022, Proceedings, Part I},
  pages={618--629},
  year={2023},
  organization={Springer}
}