Image Classification Feature Filtering for Domain Invariance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image classification models learn domain information irrelevant to image classification, leading to low accuracy in classifying images with different styles, as they fail to distinguish between domain features and class features.
Innovation Solution
An image processing method that obtains image features including domain and class features, filters out domain features with low relevance to image classification, and determines the image class based on the class features, using a feature filtering layer in a trained model to subtract domain features from image features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the image classification model learns from training samples including domain information, then the model can process various types of images, but the classification accuracy decreases when domain information is irrelevant to the classification task
Solution Approach 1:
The patent segments the image feature into two distinct components: domain feature and class feature. The domain feature captures style-specific characteristics (e.g., painting style, photograph type) while the class feature captures classification-relevant characteristics (e.g., object category). This segmentation allows the model to process diverse image domains while maintaining high classification accuracy by using only the class feature for classification decisions.
Solution Approach 2:
The patent extracts and removes the domain feature from the image feature before performing classification. By separating the domain feature (which is irrelevant to classification) from the class feature (which is relevant to classification), the model achieves high classification accuracy across different domains without being influenced by style-specific information.
2Adaptability or versatility
If the model uses all image features including domain features for classification, then the model can handle diverse image styles, but the classification accuracy drops for images with new domain information
Solution Approach 1:
The patent segments image features into domain-specific and class-specific components. The domain feature adapts to various image styles while the class feature maintains consistent classification relevance. This segmentation enables the model to reliably classify images even when encountering new domain information, as the class feature remains stable across different domains.
Solution Approach 2:
The patent extracts the domain feature and removes it from the classification process. By taking out the domain feature (which varies with image style) and using only the class feature for classification, the model achieves high reliability when processing images with new or unseen domain information.
3Adaptability or versatility
If the model learns domain features from training data, then the model can recognize images within the training domains, but the model fails to accurately classify images with unseen domain styles
Solution Approach 1:
The patent segments the image feature into domain feature and class feature, where the class feature is trained to be domain-invariant. This segmentation allows the model to achieve high classification accuracy for unseen domains by relying on the domain-invariant class feature rather than domain-specific features that would fail to generalize.
Solution Approach 2:
The patent extracts and removes the domain feature from the classification process. By taking out the domain feature (which is specific to training domains) and using only the class feature for classification, the model achieves high measurement precision even for images with unseen domain styles.
Data Source
AI summary
The present disclosure provides an image processing method, an electronic device, and a computer-readable storage medium. The method comprises: obtaining an image feature corresponding to a first image, where the image feature comprises a domain feature and a class feature; performing a feature filtering processing on the domain feature in the image feature to obtain the class feature of the first image, where a relevance between the domain feature and an image classification of the first image is below a first threshold; and determining an image class of the first image according to the class feature.