Target Recognition via Constrained Segmentation and Model Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automatic target recognition systems are limited by signature variability and dependence on phenomenological prediction technology, making them impractical for military and other applications where targets attempt to avoid recognition.
Innovation Solution
The system employs constrained image segmentation using target geometry models and a hierarchical adaptive region model matcher, incorporating a Hausdorff Silhouette Matching process and Minimum Description Length evaluation to generate and verify target hypotheses, independent of specific features and objective functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If statistical systems are trained to recognize targets based on empirical target data, then recognition can perform adequately under highly constrained conditions, but the system becomes less robust with respect to signature variability
Solution Approach 1:
The patent applies segmentation by dividing the target recognition problem into multiple hypotheses about target presence and characteristics. The system segments the image analysis process into hypothesis generation, verification, and scoring stages, allowing it to handle signature variability by evaluating multiple possible target configurations simultaneously rather than relying on a single empirical template
Solution Approach 2:
The patent changes parameters by using a hypothesis-based approach that varies target location, orientation, and presence status as adjustable parameters. The verification process evaluates multiple parameter combinations (hypotheses) about target characteristics, allowing the system to adapt to signature variability by finding the best-matching parameter set rather than requiring fixed empirical templates
2Reliability
If model-based systems use synthetic predictions of target signatures to make recognition decisions, then the system displays consistent performance across varying scene conditions, but it is subject to limitations in current signature prediction technology
Solution Approach 1:
The patent introduces an intermediary hypothesis verification process between the model-based predictions and final recognition decisions. Rather than directly relying on synthetic signature predictions, the system uses hypotheses as intermediaries that can be verified against actual image data, combining the consistency benefits of model-based approaches with the accuracy benefits of data-driven verification
Solution Approach 2:
The patent implements feedback through the hypothesis verification process, where initial model-based predictions are tested against actual image features, and the results feed back into refining the hypotheses. This feedback loop allows the system to maintain consistency across varying conditions while improving accuracy through iterative verification and adjustment of target hypotheses
3Adaptability or versatility
If optical imagery is used for target recognition, then the system can operate in various spectral regions, but the imagery is highly variable and inconsistent, making automatic recognition challenging
Solution Approach 1:
The patent applies universality by creating a hypothesis-based recognition framework that can operate across multiple spectral regions and imaging modalities. The hypothesis verification process is designed to be modality-agnostic, allowing the same fundamental approach to work with different types of optical imagery while maintaining consistent performance across varying spectral conditions
Data Source
AI summary
A target recognition system and method. The inventive method includes the steps of first receiving an image of a scene within which a target may be included and second using constrained image segmentation as a basis for target recognition. In the specific embodiment, the segmentation is variable and derived from target geometry models. The constrained image segmentation is adapted to use target geometry models to derive alternative image segmentation hypotheses. The hypotheses are verified using a novel hierarchical adaptive region model matcher. In the best mode, the hierarchical adaptive region model matcher is fully hierarchical and includes the steps of receiving an image and a hypothesis with respect to a target in the image. In the illustrative embodiment, the hypothesis contains the type, location, and/or orientation of a hypothesized target in the image. The hypothesis is received by a model tree generator. A target model library provides a plurality of computer generated target models. The model tree generator retrieves a model from the library based on a hypothesized target type and renders a model tree therefrom that represents the appearance of a hypothesized target at a hypothesized location and orientation. In the illustrative embodiment, the verification model further includes a search manager adapted to receive a model tree from the model tree generator. A feature extractor is included along with a segmentation evaluator. The segmentation evaluator is coupled to receive image features from the feature extractor. The segmentation evaluator is adapted to output a segmentation score to the search manager with respect to a specified set of segmentation labels based on a supplied set of image features for an active node.