Implicit Feature Alignment: Learn to Convert Text Recognizer to Text Spotter

This is a pytorch-based implementation for paper Implicit Feature Alignment: Learn to Convert Text Recognizer to Text Spotter.

Due to the company's code confidentiality requirements, we only release the code of ExCTC on the IAM dataset.

Personally, I dont't think it is a thorough work, but I hope this idea is useful.

Requirements

• Python3
• PyTorch1.6
• TorchVision
• OpenCV
• PIL (Pillow)
• edit_distance
• Numpy

Data Preparation

We re-crop the lines from IAM full-page images. This operation is to remove the edge phenomenon in the official line images.

Official image:

Re-cropped image:

These re-cropped images are used as training set.

Training

Trained parameter: https://drive.google.com/drive/folders/1rXJ9at9erPN6v5nPAOGWsMWFUiwBq8D6?usp=sharing

Crop the training images (data/crop_images.py) and modify the path in configuration files (cfgs.py). Then

	python main.py

A simple demo:

Ackowledgement

We use the augmentation toolkit released by RubanSeven to train the network.