Object Recognition Mask Updates for Low-Data Accuracy Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object recognition methods require elaborate artificial intelligence models and large training datasets, often resulting in inaccurate results due to complexity and resource constraints, especially in embedded environments with limited data availability.
Innovation Solution
A method using a Markov random field (MRF) model for test-time optimization, refining object recognition results by updating mask heads based on objective functions, without the need for additional training data or model re-training, to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional object recognition methods use elaborate AI models and large training datasets, then recognition accuracy can be improved, but device complexity and resource requirements increase
Solution Approach 1:
The patent segments the object recognition process into two distinct phases: (1) an initial recognition phase using a trained AI model to generate preliminary results, and (2) a refinement phase using test-time optimization to improve accuracy. This segmentation allows the system to use a relatively simple initial model while achieving high accuracy through the subsequent refinement process, thereby reducing overall model complexity requirements.
Solution Approach 2:
The patent performs preliminary object recognition using a trained AI model before applying test-time optimization. This preliminary action generates initial mask results that serve as input for the refinement process, allowing the system to start with a simple model and progressively improve accuracy without requiring an overly complex initial model.
2Measurement precision
If conventional object recognition methods use large training datasets, then recognition accuracy can be improved, but loss of time and resources increase
Solution Approach 1:
The patent performs a single training phase beforehand to obtain a basic AI model, then uses test-time optimization during inference to achieve high accuracy without requiring additional training. This preliminary action consolidates all necessary learning into one training event, eliminating the need for continuous retraining with large datasets and significantly reducing time losses.
Solution Approach 2:
The test-time optimization process is self-service in nature, automatically refining recognition results using the already-trained model without requiring external training data or additional training time. The system serves itself by improving its own output quality through the optimization process, eliminating the need for external training resources.
3Device complexity
If conventional object recognition methods are applied in embedded environments with limited data, then device complexity is reduced, but recognition accuracy deteriorates
Solution Approach 1:
The patent segments the recognition system into a simple initial model suitable for embedded environments and a refinement component that operates at test time. This allows the embedded device to use a lightweight model while achieving high accuracy through the segmentation-based refinement process that separates initial recognition from accuracy improvement.
Solution Approach 2:
The patent changes parameters during test-time optimization by adjusting mask thresholds and refinement parameters based on the specific input image characteristics. This allows the system to adapt to different conditions without requiring retraining, maintaining high accuracy in embedded environments with limited data by dynamically adjusting parameters rather than relying on complex pre-trained models.
Data Source
AI summary
A method of recognizing an object based on mask updates is provided. The method includes generating first result information in which a foreground and background of a predetermined input image have been separated from each other, generating second result information in which the results of the object recognition within the input image and a boundary portion of the object have been matched, calculating an objective function to be applied to each mask head based on the first and second result information and parameter information for the first and second result information, updating all of the mask heads based on the objective function, and performing object recognition based on the updated mask.