Image Classification using Frequent Subgraph Mining of Planar Graphs

This work details research on both generating planar graphs for natural images (Caltech-256 dataset), and using machine learning on frequent subgraphs of those images to recognise common objects.

This work includes:

• libopenpg: A python library for creating / manipulating open planar graphs.
• image-pipeline : A python program for processing images, converting them into open planar graphs, and then looking for frequent subgraphs to determine objects.

The theory behind the idea is interesting, but the observed results, while better than random guessing, were significantly worse than state of the art.

Update 2018 Other than the gSpan and libopenpg repositories (and a single paper for a class) this line of research didn’t lead anywhere and I’m concentrating on other areas now. Deep learning vastly outperforms this approach.