Learning-Based Image Page Index Selection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing image search engines fail to effectively index and return the most informative images and associated pages due to differences in objectives between image and web search, leading to omission of valuable image pages with high webpage static rank not containing many images, and pages with interesting images not being recognized as important.

Innovation Solution

A learning-based framework is employed to build a statistical model using user click data to predict the image page static rank, incorporating features like link-based, click, and URL pattern features, and aggregating image clicks over URL segments and site models to improve image page index selection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If all webpages are indexed for image search, then complete coverage of image pages is achieved, but network bandwidth, data storage, and data processing resources are overwhelmed

Engineering Contradiction:

Improvecoverage of image pagesVSAvoiddata storage and processing capacity

Core Design Contradiction:

Quantity of substanceVSDevice complexity

Solution Approach 1:

The patent segments the webpage indexing process by introducing a two-stage approach: first discovering URLs through web crawling, then selectively indexing only those URLs that contain images and meet quality thresholds. This segmentation divides the overwhelming task of indexing all webpages into manageable stages, reducing the burden on storage and processing resources while maintaining coverage of relevant image pages.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Instead of indexing all discovered URLs (excessive action), the patent applies partial indexing by filtering URLs through image detection and quality assessment mechanisms. Only URLs that contain images meeting certain quality criteria are indexed, thereby reducing storage requirements and processing complexity while ensuring that the indexed subset is highly relevant to image search queries.

Inventive Principle:

Principle #16Partial or excessive action

2Reliability

If webpage static rank is used for index selection, then web search importance is maintained, but image search relevance is lost due to omission of pages with interesting images but low static rank

Engineering Contradiction:

Improveweb search importanceVSAvoidimage search relevance

Core Design Contradiction:

ReliabilityVSLoss of information

Solution Approach 1:

The patent merges multiple ranking signals into a unified page importance score. Instead of relying solely on traditional webpage static rank, the system combines static rank with dynamic image-based features such as image quality metrics, click-through rates, and visual engagement signals. This merged approach ensures that pages with interesting images but low static rank can still be identified and indexed, maintaining both web search importance and image search relevance.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The patent incorporates feedback mechanisms where user interactions with image search results (clicks, dwell time, etc.) are used to refine page importance scores. This feedback loop allows the system to learn which pages contain valuable images and adjust indexing priorities accordingly, ensuring that pages with interesting images are not omitted even if their initial static rank was low.

Inventive Principle:

Principle #23Feedback

3Adaptability or versatility

If image search objectives are differentiated from web search, then image-specific relevance is improved, but the amount of valuable image pages omitted increases

Engineering Contradiction:

Improveimage search specializationVSAvoidnumber of indexed image pages

Core Design Contradiction:

Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent applies local quality assessment by evaluating different aspects of webpage quality separately. Instead of using a single global ranking metric, the system assesses image quality, text quality, and overall page quality independently, then combines these local assessments to determine indexing priority. This allows the system to identify and index pages with high-quality images even if other aspects of the page are mediocre, thereby maintaining image search specialization while reducing omissions of valuable image pages.

Inventive Principle:

Principle #3Local quality

Data Source

PatentUS9070046B2Learning-based image webpage index selection

Publication Date: 2015.06.30 MICROSOFT TECHNOLOGY LICENSING LLC

AI summary

Architecture that performs image page index selection. A learning-based framework learns a statistical model based on the hyperlink (URL-uniform resource locator) previous click information obtained from the image search users. The learned model can combine the features of a newly discovered URL to predict the possibility of the newly-discovered URL being clicked in the future image search. In addition to existing web index selection features, image clicks are added as features, and the image clicks are aggregated over different URL segments, as well as the site modeling pattern trees to reduce the sparse problem of the image click information.