Building a robust classification pipeline using clustering centroids to categorize unseen images, filter out anomalies with confidence thresholds, and benchmark against zero-shot CLIP.
