This thesis aims to automatically organize a collection of images in order to identify recurring patterns related to brand identity. The work builds on the FRESCO framework, which translates concepts from visual semiotics into computational descriptors across three levels (plastic, figurative, and enunciational) and produces an image-to-image similarity measure, the FRESCO-score. Starting from these similarity measures, the goal is to develop an unsupervised clustering pipeline capable of generating coherent and interpretable groups while reducing the impact of noise and anomalous cases. Before applying the clustering algorithms, cleaning and filtering steps are implemented on similarity relations, which are represented as an affinity graph.