Digital Image Decomposition via Heuristic Search and Zone Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional digital image decomposition methods rely on brittle ad hoc rules and specific visual aspects, making them impractical for large volumes of documents and inconsistent in results.
Innovation Solution
A system and method using heuristic search to decompose digital images into zones based on generic visual features, employing a learned generative zone model with likelihood and prior models, and binary integer linear programming or A* best-first graph search to infer an optimal set of non-overlapping zones.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional bottom-up or top-down decomposition methodologies are used, then document images can be broken down into zones, but the results are inconsistent and brittle due to reliance on ad hoc rules
Solution Approach 1:
The patent replaces the mechanical rule-based system with a machine learning-based system. Instead of using hand-crafted ad hoc rules for decomposition, the system employs trained classifiers and probabilistic models that automatically learn decomposition patterns from training data, thereby improving consistency while reducing the brittleness associated with rigid rule-based approaches
Solution Approach 2:
The patent transforms the decomposition problem from a rule-based decision process into a probabilistic parameter optimization problem. By using likelihood models and prior models with associated cost functions, the system changes the approach from binary rule application to continuous parameter optimization through Viterbi decoding, improving reliability through statistical rather than deterministic methods
2Measurement precision
If manual document decomposition is performed, then accurate zone identification is possible, but it becomes impracticable for large volumes of documents
Solution Approach 1:
The patent implements a self-service system where the machine learning model automatically performs document decomposition without human intervention. The trained classifiers and probabilistic models process documents autonomously, maintaining high accuracy while achieving the scalability needed for large-volume document processing
Solution Approach 2:
The patent extracts the decomposition expertise from manual operators and encodes it into trained machine learning models. By transferring the knowledge embedded in manual decomposition practices into automated algorithms, the system preserves accuracy while eliminating the productivity limitations of manual processing
3Ease of manufacture
If ad hoc rules are used for decomposition, then implementation is straightforward, but the system produces varying results even with little actual change in data
Solution Approach 1:
The patent replaces fragile ad hoc rules with robust machine learning models that have been trained on comprehensive data. These models generalize better to variations in document layouts and content, providing stable results even when input data varies slightly, while still maintaining relatively straightforward implementation through standard ML pipelines
Data Source
AI summary
A system and method for decomposing a digital image is provided. A digital image is represented as a word-graph, which includes words and visualized features, and zone hypotheses that group one or more of the words. Causal dependencies of the zone hypotheses are expressed through a learned generative zone model to which costs and constraints are assigned. An optimal set of the zone hypotheses are inferred, which are non-overlapping, through a heuristic search of the costs and constraints.