Image Search Query Expansion via Metadata Term Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image search technologies face challenges in returning both relevant and diverse results due to the limited expressiveness of keyword-based query formulation and the presence of word-sense and type-specific ambiguities, especially with images lacking adequate annotations and noisy data.
Innovation Solution
The method involves generating an updated query by extracting terms from metadata associated with selected search results, weighting them based on frequency, and re-ranking to enhance topical diversity and relevance, ensuring that the results set reflects multiple interpretations of ambiguous queries while maintaining precision for unambiguous ones.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If keyword-based query formulation is used for image search, then search simplicity is maintained, but search result diversity deteriorates due to word-sense and type-specific ambiguities
Solution Approach 1:
The patent segments the search result generation process into multiple independent components: (1) executing the original query to get initial results, (2) extracting terms from metadata of selected results, (3) generating an updated query with extracted terms, and (4) executing the updated query to get additional results. This segmentation allows the system to maintain simple keyword-based querying while incorporating diverse result sources through separate processing stages.
Solution Approach 2:
The patent introduces an intermediary process between the original query and final results: term extraction from metadata of initial search results. This intermediary step generates additional search terms that capture different interpretations and contexts, which are then used to expand the search query. This mediator resolves the contradiction by bridging simple keyword search with diverse result generation without requiring complex user input.
2Measurement precision
If traditional information retrieval models are used, then relevance is optimized, but diversity is lost by not considering multiple interpretations of ambiguous queries
Solution Approach 1:
The patent performs preliminary action by executing the original query first to obtain an initial set of search results. From these results, it extracts terms from metadata that represent different interpretations and contexts. This preliminary extraction of diverse terms before final result generation ensures that multiple interpretations are captured while maintaining relevance through the original query execution.
Solution Approach 2:
The patent adds another dimension to the search process by incorporating metadata-derived terms from initial results into an updated query. This creates a two-dimensional search approach: the original query dimension for relevance and the metadata-derived terms dimension for diversity. By executing both queries and combining results, the system achieves both relevance accuracy and interpretation coverage simultaneously.
3Measurement precision
If image search returns only the most relevant results, then precision is improved, but comprehensiveness deteriorates by missing alternative interpretations
Solution Approach 1:
The patent segments the result set into two distinct components: (1) results from the original query that ensure precision and relevance, and (2) results from the updated query with metadata-derived terms that ensure comprehensiveness and diversity. By presenting both segments to the user, the system maintains precision while simultaneously improving comprehensiveness without mixing the two result types.
Solution Approach 2:
The patent merges two separate search executions: the original query results and the updated query results. By combining these two result sets, the system achieves both precision (from the original query) and comprehensiveness (from the updated query with additional terms). This merging allows the system to deliver a comprehensive result set that includes both the most relevant results and alternative interpretations.
Data Source
AI summary
Techniques are described to increase the diversity or focus of image search results. A user submits an original query to search for images. A server generates a first results set by executing the original query using metadata associated with each image. The server selects, from the first results set, a specified number of results ranked highest and generates a list of terms from the metadata of each of the results selected. The terms may be only the tags of the results. The server generates an updated query using terms in the list that may be weighted based on the frequency of the term in the list or include only a specified number of the highest occurring terms in the list. The server generates a second results set by executing the updated query using metadata associated with each image. The second results set is then stored and displayed to the user.