Supporting code to the publication “The cracks that wanted to be a graph”: application of image processing and Graph Neural Networks to the description of craquelure patterns published in ICCV 2019 Workshops.

Taking a skeletonized binary image as an input, given algorithm extracts non-directed graph from a cracks pattern, classifies nodes by topology onto X, Y, and O types, fits edges with polynomial, and exports comprehensive characteristic of a craquelure pattern. The latter can be used for forgery detection, origin examination, aging monitoring, and damage identification.

We thank alchemyst and phi-max for their algorithms which we modify and apply in our code. All the rights to original implementations belong to the authors.

The code is written under MatLab R2017b, other versions haven't be tested. Unlikely anything but the Image Processing Toolbox is required. If you find any surprising dependency - please notify us.

Binarization of the crack images is very tricky and ungreatful process so we leave it for a user's responsibility. (Although we provide a helper code prepare_bw.m which was used in our experiments (parameters are in filenames of the images)).

Run main.m for a fast start.

The algorithm takes a bunch of labeled graphs, uses them to train GNN, and then extracts a vector of hidden features from the GNN's layers for each graph.

The implemention is based on this algorithm by Xu et al.
Requirements:

pytorch
tqdm
numpy
networkx
scipy

As simple as:

cd GNN
python main.py --dataset CRACKS --lr 0.001 --epochs 10 --fold_idx 0

The output is a .mat file graph-features.mat containing only one variable of a size [N_graphs x N_features].

If you want to use custom dataset run createTXT.m and move the output to /dataset/CRACKS/.

The structure of the required .txt is following:

• each graph is a block
• first line of a block consist of %number of nodes% %class label%
• each following line describes single node in a way: %node label% %number of connected nodes% %connected node #1% %connected node #2% %connected node #3%...
• row number correspond to the node's index, starting from 0
• test/train partition is defined by cross-validation and doesn't appear in .txt

For example:

10 7
0 3 1 2 9
0 3 0 2 9
0 4 0 1 3 9
0 3 2 4 5
0 3 3 5 6
0 5 3 4 6 7 8
0 4 4 5 7 8
0 3 5 6 8
0 3 5 6 7
1 3 0 1 2

The block corespond to a graph which consist of 10 nodes and belongs to class 7. First (0) node has label 0 and has 3 neighbours; these neighbours are nodes 1, 2, and 9. The same can be applied to the next nodes.

The final part where extracted features are combined and used for the classification of patterns.

1. Use Option 2 to generate graph-features in advance or load proposed stats.m and graph-features-263.m
2. Follow classification.m

Please cite the paper if you find our algorithm useful.