Coupled Learning Model for Cross-Domain Image Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing identification technologies struggle to perform accurate learning and inference when handling images with different domains, such as live-action RGB images and thermal infrared images, due to significant differences in features like color, illumination, and frequency components.
Innovation Solution
A learning device utilizing a coupled mathematical model that includes a shallow CNN to generate domain-shared features and a deep CNN for inference, employing domain relaxation learning to adjust and emphasize features like color, illumination, low frequency components, and high frequency components during learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional machine learning methods are used to handle images with different domains, then the model can process diverse image types, but the learning and inference cannot be performed correctly due to large feature differences
Solution Approach 1:
The patent segments the feature extraction process into domain-specific feature extractors and domain-shared feature extractors. Each domain (e.g., RGB, TIR) has its own feature extractor that captures domain-specific characteristics, while a shared extractor learns common features across domains. This segmentation allows the model to handle diverse image domains while maintaining learning accuracy by processing each domain's unique characteristics separately before combining them.
Solution Approach 2:
The patent introduces domain-shared features as an intermediary representation that bridges different image domains. The domain-shared feature extractor learns features that are common across multiple domains (such as structural patterns, shapes, or semantic content) and uses these as a mediator to connect domain-specific features. This intermediary layer enables the model to transfer knowledge between domains while adapting to domain-specific variations.
2Measurement precision
If domain-specific features are emphasized to maintain accuracy for each domain, then learning accuracy improves, but the ability to generalize across domains deteriorates
Solution Approach 1:
The patent merges domain-specific features and domain-shared features in the feature fusion layer. The domain-specific feature extractors capture unique characteristics of each image type (e.g., color information for RGB, thermal patterns for TIR), while the domain-shared feature extractor captures commonalities. These feature sets are then combined through concatenation or addition, allowing the model to leverage both specialized knowledge and general patterns for improved performance across all domains.
Solution Approach 2:
The patent designs the domain-shared feature extractor to serve multiple domains simultaneously, making it a universal component that can process features from any input domain. This multi-functional extractor learns representations that are applicable across different image types, enabling the model to generalize well to unseen domains while maintaining the ability to capture domain-specific nuances through the dedicated feature extractors.
3Speed
If a simple mathematical model is used for faster learning, then learning speed improves, but the ability to handle complex cross-domain features deteriorates
Solution Approach 1:
The patent segments the learning process into two stages: first learning domain-shared features that are common across all domains, then learning domain-specific features. This segmentation allows the model to quickly capture general patterns early in training (fast learning) while progressively learning more complex domain-specific characteristics. The staged approach prevents the model from being overwhelmed by the complexity of all domains simultaneously, improving both learning speed and cross-domain capability.
Solution Approach 2:
The patent performs preliminary learning of domain-shared features before fine-tuning domain-specific features. By first establishing a foundation of common features that apply across all domains, the model creates a robust baseline that accelerates subsequent learning. This preliminary action ensures that the model has already captured essential cross-domain patterns before tackling domain-specific complexities, improving both learning efficiency and adaptability.
Data Source
AI summary
A learning device according to the present disclosed technology is a learning device including a coupled mathematical model capable of machine learning and learning a data set of a target domain from a data set of an original domain for a teacher, in which a pre-stage part of the coupled mathematical model generates a plurality of low-level feature maps from input image data, compares the low-level feature maps of data sets belonging to the same type of learning target for the original domain and the target domain in the image data, and calculates domain-shared features, and calculates domain relaxation learning information for each space of {1} color, {2} illumination, {3} low frequency component, and {4} high frequency component among the domain-shared features.