Image clustering for unsupervised analysis of plankton data

Abstract

Advancements in automated imaging has made it possible to enhance the data both in terms of quantity and quality. This has prompted the development of plankton imaging systems for acquiring the species level information of the communities. However, screening the huge amount of data has been a challenge for both humans and computers. This Master’s thesis project focused on visual clustering of plankton image data, by implementing and applying image clustering methods on plankton image data sets. These data set were collected from the Baltic Sea using an imaging flow cytometer. In order to form the clusters, first the features were extracted using AlexNet and ResNet-18. The extracted features of each CNN were clustered using the Hierarchical and K-means algorithms and evaluated by checking the purity of the clustering result. The results showed that ResNet-18 is better in feature extraction and that for a small number of classes the K-means method has the highest level of purity. However, the Hierarchical method shows higher purity in case the number of clusters is low. On the other hand, the Hierarchical method shows better purity when the number of classes is large.