Image Search Query Representation via Click Log Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image search technologies face limitations in accurately describing image content due to the mismatch between text queries and image formats, with existing methods relying on text matching, classification, and click-through rates, which are either inaccurate, semantically deviated, or biased.
Innovation Solution
Generating an image characteristic representation by clustering image characteristic vectors from clicked images based on their content, allowing for direct correlation calculation between query and image, thereby improving search results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If text matching characteristic is used for image search correlation, then the search process is simple, but the accuracy is affected because surrounding text may be inconsistent with image content
Solution Approach 1:
The patent introduces an intermediary text generation process that extracts actual content from images through OCR and other recognition technologies, then uses this extracted text as the basis for correlation calculation. This intermediary step ensures the text truly represents image content while maintaining the simplicity of text-based search operations.
2Measurement precision
If classification matching characteristic is used with fine categorizing system, then classification accuracy decreases, but matching with query becomes more difficult due to semantic deviation
Solution Approach 1:
The patent extracts the most relevant classification labels that directly correspond to query semantics, rather than using the entire fine-grained classification hierarchy. This extraction approach maintains high classification accuracy while reducing semantic deviation and matching difficulty by focusing only on the most pertinent categories.
3Reliability
If click-through rate characteristic is used for correlation measurement, then user behavior data is leveraged, but the data has biases and noises and is sparse for many queries
Solution Approach 1:
The patent combines multiple correlation characteristics (text matching, classification matching, and click-through rate) into a unified correlation calculation framework. This merging approach allows the system to leverage user behavior data where available while compensating for its sparsity and noise through the other characteristics, thereby improving overall measurement precision and reliability.
Data Source
AI summary
Embodiments of the present invention disclose a method and apparatus of generating an image characteristic representation of a query, and an image search method and apparatus. The method of generating an image characteristic representation of a query comprises: acquiring a clicked image set corresponding to a target query based on an image click log of a search user; generating image characteristic vectors corresponding to clicked images in the clicked image set based on image content characteristics of the clicked images; and clustering the clicked images based on the image characteristic vectors, and using a clustering result as an image characteristic representation of the target query.