Feature Extraction Network Training with Two-Stage Class Center Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image feature extraction networks trained with a mix of high-quality and low-quality sample images face challenges due to class center distribution offsets, affecting training effectiveness.
Innovation Solution
A two-stage training method is employed, where initial class center matrices are adjusted based on first and second losses to align low-quality sample features with main class centers, improving feature extraction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a feature extraction network is trained with a mix of high-quality and low-quality sample images, then the network can handle diverse input conditions, but class center distribution offsets occur affecting training effectiveness
Solution Approach 1:
The patent segments the training process into two distinct stages: a first training stage that processes all samples uniformly, and a second training stage that specifically addresses low-quality samples. This segmentation allows the network to first learn general features from diverse data, then refine its ability to handle challenging inputs without compromising overall training effectiveness.
Solution Approach 2:
The patent performs preliminary action by establishing initial class centers during the first training stage before addressing the specific problem of low-quality samples. These pre-established class centers serve as reference points that guide the second training stage, preventing class center distribution offsets from occurring in the first place.
2Measurement precision
If multiple class centers are maintained for each class, then fine-grained classification is possible, but class center offset occurs when training with difficult samples
Solution Approach 1:
The patent performs preliminary action by establishing initial class centers during the first training stage before addressing the specific problem of low-quality samples. These pre-established class centers serve as reference points that guide the second training stage, preventing class center distribution offsets from occurring in the first place.
Solution Approach 2:
The patent implements feedback mechanisms where the network's performance on low-quality samples during the second training stage informs adjustments to the training process. The loss function provides continuous feedback about how well features from difficult samples align with their corresponding class centers, enabling iterative refinement that maintains both precision and stability.
Data Source
AI summary
Embodiments of this application provide a feature extraction network training method, a classification method, an apparatus, and an electronic device. The method includes: adjusting initial class center matrices respectively corresponding to a plurality of class centers of each of a plurality of classes, and an image feature of the first sample image extracted from a to-be-trained feature extraction network, a parameter of the to-be-trained feature extraction network and the initial class center matrices, to obtain a preliminarily trained feature extraction network and first class center matrices respectively corresponding to the class centers; and adjusting the first class center matrices, and an image feature of the second sample image extracted from the preliminarily trained feature extraction network, a parameter of the preliminarily trained feature extraction network, to obtain a trained feature extraction network.