Newspapers_clustering_using_py-faster-rcnn

In this repository we propose a Deep Learning approach for document clustering. This is an improvment of a previous work that used traditional Machine Learning approach.

Description

We use py-faster-rcnn code from R.Ghirschik to learn text and image detection.

Dataset

The used dataset is an ensemble of 102 Russian newpapers pages annotated and taken from UCL Machine learning site. This is the dataset we use to tune faster-rcnn code (68 for training, 34 for testing)

We show below an example with ground truth:

page

ground truth

Formatting data

To use faster-rcnn code, we need to reformate dataset like Pascal-VOC challenge (comprising xml annotation files for bouding boxes).

But first of all, clone our fork of py-faster-rcnn code into your home for example. This forl contains appropriate to deal with our newspapers dataset. Then open import_data.py, fill main_path variable with following path yourhome/py-faster-rcnn/data/NewsPapers/UCL and run it :

python import_data.py

It will create appropriate reformatting of our newspapers dataset into yourhome/py-faster-rcnn/data/NewsPapers/UCL

Training/testing

Now go to yourhome/py-faster-rcnn and run the following command line

./experiments/scripts/faster_rcnn_alt_opt.sh 0 VGG_CNN_M_1024 newspapers

You will get test results into yourhome/py-faster-rcnn/data/NewsPapers/UCL/results, containing bounding boxes for text and for illustration, for each test image present into yourhome/py-faster-rcnn/data/NewsPapers/UCL/ImageSets/Main/test.txt

Results

We present here after some testing results for text/illustration

example

text

illustration

example

text

illustration

example

text

illustration