Hi everyone. I'm currently trying to assess the ability of human volunteers to cluster a set of images into a fixed number of clusters according to perceived visual similarity (the images are self-organized maps of breast tumours' gene expression). In order to do that I build an averaged identity matrix (with binarized similarity: 1 is belonging to the same group, and 0 not belonging), which I later transform into a graph and perform the partition. I'm stuck in the part in which I contrast human ability to an algorithm. I was wondering if there's an algorithm or software which transforms a set of images into an undirected wighted graph, where the weight of the edges represents the similarity between images (nodes). This would be my best scenario, because I would be able to perform the same partition method and compare. Else, I'd still appreciate some suggestions for image clustering according to visual similarity. Thanks in advance.
Daniel,
I could find a direct healthcare example but here is public data which transforms imagines. Hope this script helps!
Best,
John
Hi,
To cluster a set of images, first of all, you will select a way to represent an image (SURF, SIFT, GIST, ... you can read following paper for more details http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.366.6730&rep=rep1&type=pdf) and then apply an appropriate clustering method (i suggest GHSOM - Growing Hierarchical Self Organizing Map or K-means for this case).
Regards,
Dinh
In a first step, create a graph with undirected, unweighted edges. An edge means similarity. Then, in order to obtain a weighted edge between two vertices, determine how many neighbors belong to both vertices, divided by the total sum of neighbors of these two vertices.
For automatic computation of similarity there are lots of options. One is to downsize the image, to say, 16*16 pixels, then comparisons are between vectors with 256 components. The method of choice strongly depends on the type of images.
Regards,
Joachim
Actually, it depends on the size of similarity matrix. when you are trying to cluster the phenotypes inside of the images, your similarity matrix is not following the Euclidean distance , and normally the distance are geodesic. So, the traditional methods like K-Means are not applicable for that. If the number of clusters are small enough , you can go a head and use SPECTRAL CLUSTERING with an RBF kernel, otherwise, you should consider that in spectral clustering with a large dimension , exhaustive O(n2) pairwise comparison between images is computationally intractable. In this case you can go a head and test other methods like t-SNE or UMAP. But, considering your explanation, it seems that for your specific application SPECTRAL CLUSTERING works well.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Bioinformatics