Automated design of feature extraction for unsupervised image clustering using grammatical evolution
Mia Gerber, Nelishia Pillay · 2022 IEEE Symposium Series on Computational Intelligence (SSCI) · 2022
When doing image processing it has been shown that the use of an appropriate feature extraction function can result in a significant boost in performance. Selecting or creating the correct feature extraction function is however non-trivial. This work uses grammatical evolution (GE) to evolve a feature extraction function for images. The goal is to evolve feature extraction functions that are able to improve the clustering results of three different unsupervised clustering algorithms: KMeans++, DBSCAN and Mean Shift. In order to test whether the feature extraction functions evolved by the GE are effective, the clustering algorithms are also used with no feature extraction and with randomly generated feature extraction functions. The results of experimentation show that using a GE to evolve a feature extraction function is more effective than using no feature extraction or randomly generated feature extraction, with the exception of K-Means++ when applied to the Oral lesion dataset where GE feature extraction performed equally as good as when no feature extraction is used. A possible reason for this might be that the grammar for the GE needs additional operators to generate color specific features. Future work includes applying transfer learning to the GE.