wangweilai1/textfcn

A Caffe Fully Convolutional Network that detects any kind of text and generates pixel-level heatmaps.

Text Detection FCN

A Caffe Fully Convolutional Network that detects any kind of text and generates pixel-level heatmaps. Adapted from Fully Convolutional Models for Semantic Segmentation by Jonathan Long, Evan Shelhamer, and Trevor Darrell. CVPR 2015 and PAMI 2016.

Model

A COCO-Text trained model is available here.

Publications

This FCN was used to improve the TextProposals algorithm by Lluis Gomez. The improved version is available here. That lead to two publications, which you may cite if using this FCN:

FAST: Facilitated and Accurate Scene Text Proposals through FCN Guided Pruning
Dena Bazazian, Raul Gomez, Anguelos Nicolaou, Lluis Gomez, Dimosthenis Karatzas, Andrew D.Bagdanov.
Pattern Recognition Letters, 2017.

Improving Text Proposals for Scene Images with Fully Convolutional Networks
Dena Bazazian, Raul Gomez, Anguelos Nicolaou, Lluis Gomez, Dimosthenis Karatzas and Andrew Bagdanov.
ICPR workshop (DLPR), 2016.

Requirements

The amount of required GPU memory depends on the image size:

1000x1000 --> 2.8 GB
512x512 --> 1.6 GB

It takes 0.17s per image on a TitanX.

Development

This FCN was trained during MS’s thesis. Extra information about the training and usage can be found in the thesis or in this slides.