A Word Spotting Method in Scene Images based on Character Recognition
The code is containing a CAFFE model for training an FCN to recognize characters in the scene images and an application of it for word-spotting.
An FCN training for character detection (38 classes- 26 english letters incase sensetive, 10 numbers and signs)
The set up for training the net based on the CAFFE framework is at caffe/net_train/synth-fcn8s-atonce/, and solv.py is for training.
The CAFFE trained model of character recognition is at caffe/snapshot/.
In order to get the heatmaps, there are three python files at caffe/get_heatmap/. The getHeatmaps.py is for visualizing the heatmaps of a list of images. getHeatmapsSINGLE.py is for visualizing the heatmap of a single image. saveHM.py is for saving the heatmaps of the images as png files.
To try wordspotting based on these character recognition, there are some quantitative and qualitative files at caffe/word_spotting/.
Please cite this work in your publications if it helps your research:
@InProceedings{Bazazian18,
author = {D.~Bazazian and D.~Karatzas and A.~Bagdanov},
title = {Word Spotting in Scene Images based on Character Recognition },
booktitle = {Proceeding of Computer Vision and Pattern Recognition Workshops},
publisher = {IEEE},
pages = {1872--1874},
year = {2018}
}