Document Layout Detection

This dataset has been created primarily for the evaluation of layout analysis (physical and logical) methods. It contains realistic documents with a wide variety of layouts, reflecting the various challenges in layout analysis.

Dataset has four classes:

• Text Region
• Seperation Region
• Noise Region
• Image Region

Format of Dataset

Prerequisites

• Python3 - Python version 3.6
• CUDA - CUDA version 9.0

Pipeline

Run the requirements.txt

cat requirements.txt | xargs -n 1 -L 1 pip install

Convert to COCO format

Load Dataset

gtf.Train_Dataset(root_dir="../sample_dataset", coco_dir="PRImA Layout Analysis Dataset", img_dir="Images", set_dir="Train", batch_size=8, image_size=512, use_gpu=True)

Load Model

gtf.Model()

Set Hyper Parameters

gtf.Set_Hyperparams(lr=0.0001, val_interval=1, es_min_delta=0.0, es_patience=0)

To View Loss Plots (Optional)

logs_base_dir = "tensorboard/signatrix_efficientdet_coco"
os.makedirs(logs_base_dir, exist_ok=True)
%load_ext tensorboard
%tensorboard --logdir {logs_base_dir}

Start the training

gtf.Train(num_epochs=5, model_output_dir="trained/")

DEMO

Run this colab

Acknowledgments

• MONK_OBJECT_DETECTION
• DATASET