Image Clustering via Embedding Vectors for Search Ambiguity

The patent introduces image embeddings as an intermediary representation between keywords and images. Instead of directly matching keywords to images (which causes ambiguity), the system converts both keywords and images into embedding vectors in a shared semantic space, where similarity is measured through vector distance. This intermediary embedding layer resolves the contradiction by enabling fast retrieval while maintaining accuracy through semantic similarity rather than exact keyword matching.

The system changes the parameter space from discrete keyword matching to continuous embedding vector comparison. By transforming images and keywords into continuous vector representations with varying dimensions and similarity thresholds, the system can adjust the balance between speed and accuracy by modifying embedding dimensions, similarity thresholds, and retrieval parameters without fundamentally changing the search mechanism.

The patent segments diverse search results into multiple organized clusters based on embedding similarity. Instead of presenting a flat list of diverse results that are hard to navigate, the system groups results into thematic clusters (e.g., different interpretations of a ambiguous query) and presents them as organized categories. This segmentation maintains diversity while improving navigability by allowing users to explore different result groups systematically.

The system adds a new dimension to result presentation by organizing diverse images along the embedding similarity dimension. Rather than displaying diverse results in a single flat dimension (list view), the patent uses the embedding space to create hierarchical or multi-dimensional organization, allowing users to navigate results by similarity thresholds or cluster groupings, thus making diversity manageable and navigable.