Visual Fingerprinting Segmentation for Rich Document Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing two-dimensional visual fingerprinting methods for rich document content are unforgiving to minor changes, as they 'bind together' local properties of different object types, leading to hybrid mixed-content fingerprints that require precise alignment, which is at odds with user expectations and can result in false positives when content is modified.
Innovation Solution
A method and system that classify features like text, line-art, and photo objects based on local image properties to separate content into multiple layers, allowing for independent fingerprinting of each layer type, preventing the mixing of information across categories and maintaining the purity of local appearances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If hybrid mixed-content visual fingerprinting is used to capture local properties of different object types, then the fingerprinting can represent diverse content types, but the fingerprints become sensitive to minor local changes and require precise alignment
Solution Approach 1:
The patent segments the rich document content into multiple separate layers based on object type (text layer, photograph layer, line-art layer). Each layer is fingerprinted independently, creating separate fingerprints for each content type. This segmentation prevents the mixing of different object type properties in a single fingerprint, thereby reducing sensitivity to minor local changes while maintaining the ability to represent diverse content types through multiple specialized fingerprints.
2Device complexity
If the entire page content is flattened to form a single image for fingerprinting, then the process is simple and unified, but the resulting fingerprints mix properties of different object types and lose object identity
Solution Approach 1:
Instead of flattening the entire page into a single image, the patent segments the content into multiple layers based on object type while preserving the spatial relationships. Each layer is then processed independently to create type-specific fingerprints. This approach maintains object identity within each layer while avoiding the mixing of different object type properties, thus preventing information loss.
Solution Approach 2:
The patent adds a dimensional aspect by creating multiple fingerprint layers corresponding to different object types, rather than a single flat fingerprint. This multi-dimensional fingerprinting approach preserves object identity by maintaining separate representations for text, photographs, and line-art, while still capturing the overall page structure through the combination of layers.
3Measurement precision
If mixed-content fingerprints are created from closely placed objects, then the fingerprints capture local neighborhood properties, but they bind together properties of unrelated object types requiring precise visual alignment
Solution Approach 1:
The patent segments the fingerprinting process by object type, creating separate fingerprints for text, photograph, and line-art objects. This segmentation allows each fingerprint to focus on its specific object type properties without being contaminated by unrelated object types, thereby maintaining localization accuracy while providing flexibility to object rearrangement since each layer can be independently matched.
Data Source
AI summary
Visual fingerprinting is used to provide a robust and highly effective method of finding similar content in a large document collection of rich document content composed of multiple text, line-art, and photo image objects. The visual fingerprints capture unique two-dimensional localized aspects of document appearance. The visual fingerprints are highly distinctive; fast for lookup; compact for storage requirements; and scalable to large document collections.