Advancing Image Annotation : Unsupervised clustering to minimize human input in AI training

This thesis demonstrates the feasibility of using unsupervised clustering to group microscopic bead images, thereby reducing the human effort required to create annotated datasets for training AI classifiers. The purpose is to address the challenge of efficiently generating labelled data, a significant bottleneck in machine learning workflows to identify defect beads. The study evaluated a pipeline consisting of neural network-based feature extraction, dimensionality reduction, and clustering algorithms. Multiple methods were explored, including EfficientNetB0, ResNet50, ISOMAP, UMAP, t-SNE, MDS, PCA, GMM, DBSCAN, where a pipeline combining EfficientNetB0 for feature extraction, UMAP for dimensionality reduction, and DBSCAN for clustering was selected for its superior performance. Applying the pipeline on an annotated dataset generated an accuracy of 96% and F1 score (which takes class imbalance into account) of 0.95, indicating that this approach can effectively cluster data based on the visual appearance of beads and can be used to reduce the need of manual annotation. Finally, even though the development and evaluation of the final pipeline was focused on bead identification, the unsupervised design of the pipeline suggests that its use should generalize to other application areas where annotated image sets are scarce.