Dmitrii Gurov
A Multi-Stage AI Pipeline for Image Cluster Summarization Grounded in Semiotic Metadata.
Rel. Lia Morra, Francesco Manigrasso. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
Abstract
Unsupervised clustering makes it possible to organize large collections of social-media images into coherent groups, but understanding what each group actually contains still requires manual inspection. This thesis addresses the automatic summarization of image clusters: generating a short natural-language description for every cluster in a collection. The work builds on a previous thesis in which about 35,000 Instagram images of luxury perfume and fashion brands were partitioned by three unsupervised pipelines: Gaussian Mixture Models on CLIP embeddings and DBSCAN and Spectral Clustering on the FRESCO semiotic similarity. The thesis proposes an end-to-end summarization pipeline: a small set of representative images is selected from each cluster and passed to a vision-language model, which writes a draft description; then the draft is refined by injecting cluster-level metadata extracted from the FRESCO framework, weighted with TF-IDF and validated against the images with CLIP, and rewritten where it confuses specific competing clusters; the final description is selected among the candidates by a panel of independent judges (VQAScore, FAITHSCORE, G-EVAL) on top of CLIP-based retrieval, coverage and exclusivity metrics.
Descriptions are compared with ImageSet2Text, the current state of the art for image-set description
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Informazioni aggiuntive
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
