Neural Network Training for Source-Target Domain Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification models trained using conventional methods often suffer from poor accuracy due to the distribution difference between source and target domains, leading to suboptimal classification results.
Innovation Solution
A method and apparatus for training an image classification model that utilizes neural networks to determine category prediction results for both source and target domain images, and calculates a loss based on source-domain image feature pairs with the same category but different modal tagging results, aligning image features to focus on category distinction and reduce modal distinction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional training methods are used for image classification models, then the training process is simple, but the classification accuracy is poor due to distribution difference between source and target domains
Solution Approach 1:
The patent segments the training process into multiple distinct modules: feature extraction, modal tagging, category tagging, and loss calculation. Each module handles a specific aspect of the training process, allowing for more precise control over feature alignment and category classification while managing complexity through modular design
Solution Approach 2:
The patent introduces modal tagging results as an intermediary element that bridges source and target domains. By using modal tags as an intermediate representation, the system can align features across different domains before final category classification, improving accuracy without requiring direct complex mappings
2Measurement precision
If image features focus on modal distinction, then modal classification is accurate, but category differentiation is weakened leading to poor classification results
Solution Approach 1:
The patent segments feature representation into modal-specific and category-specific components. By separating these functions, the system can maintain modal information through dedicated modal tagging while directing category differentiation through category tagging, preventing information loss in either domain
Solution Approach 2:
The patent applies different quality requirements to different parts of the feature representation. Modal tagging preserves local modal characteristics while category tagging focuses on global category differentiation. This local quality approach ensures that modal information is not lost while achieving accurate category classification
Data Source
AI summary
A method for training an image classification model performed by an electronic device and includes: obtaining a plurality of sample source-domain images, a plurality of sample target-domain images, modal tagging results of the sample source-domain images, and category tagging results of the sample source-domain images; determining first category prediction results of the sample source-domain images by using a neural network model; determining first category prediction results of the sample target-domain images by using the neural network model; for a category tagging result, determining a first loss of the category tagging result based on source-domain image feature pairs corresponding to the category tagging result; and training the neural network model based on first losses of category tagging results, the first category prediction results of the sample source-domain images, and the first category prediction results of the sample target-domain images, to obtain an image classification model.