Anomaly Detection Study
To find anomalies without any supervision is not exactly an easy task.
We need to ask our computer to do this without any single label:
Source: Stanford's own result with supervised learning
However, after carefully examining a lot of supervised algorithms and a lot of different segmentation techniques, we came up with one solution that seems to works (sometimes).
To detect the anomalies:
- We first loaded the image into grayscale format:
- Then we factorise the image as a product of weight and influence by using Non negative matrix factorisation:
This image above is reconstructed by using 50 features. Reducing the features reduces the sensitivity of algorithms.
10 features reconstruction:
- Now, we analyse the weight matrix:
We can see that there are some anomalies at the bottom. We find these anomalies(or outliers) by using the DBSCAN algorithm.
Similarly, we analysed the influence matrix.
After combining both weight and influence and detecting the anomalies, we came up with the following result:
On the right, there is the original image. On the left, you can see the portions cut by the algorithm where it felt that something is not right.