Feature Fusion Model Training with Self-Attention Weight Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current sample recognition technologies face inaccuracies due to redundancy in information fusion during feature splicing, leading to unreliable sample features.
Innovation Solution
A method for feature fusion model training that involves acquiring training samples, extracting semantic and global features, splicing them to calculate autocorrelation features, and using self-attention weights to adjust and fuse features, thereby reducing redundancy and improving accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If different characterizations are spliced directly to obtain fused features, then the feature fusion process is simple and fast, but information redundancy occurs and feature accuracy decreases
Solution Approach 1:
The patent introduces an attention mechanism as an intermediary between feature extraction and fusion. The attention module computes attention weights that dynamically adjust the contribution of different characterizations during fusion, preventing simple direct splicing while maintaining computational efficiency. This intermediary layer filters redundant information and emphasizes important features.
Solution Approach 2:
The patent changes the parameters of feature fusion by introducing attention weights that dynamically adjust the importance of different characterizations. Instead of fixed equal-weight splicing, the system learns optimal weighting parameters through training, allowing flexible adjustment of feature contributions based on their relevance to the classification task.
2Measurement precision
If attention mechanism is introduced for feature fusion, then feature accuracy is improved, but model complexity increases
Solution Approach 1:
The patent segments the feature fusion process into distinct modules: feature extraction, attention weight computation, and weighted fusion. This segmentation allows the attention mechanism to be added as a separate, manageable component rather than integrating it throughout the entire model, reducing overall complexity while maintaining accuracy benefits.
3Measurement precision
If iterative model updating is performed for training, then feature fusion accuracy is improved, but training time increases
Solution Approach 1:
The patent implements continuous iterative training where the model is repeatedly updated with training samples to continuously improve feature fusion accuracy. The training process maintains momentum by using gradient descent optimization, allowing the model to converge efficiently rather than requiring complete retraining from scratch each time improvements are made.
Data Source
AI summary
A method for feature fusion model training and sample retrieval includes: inputting training sample into an initial feature fusion model to obtain a training semantic feature and a training global feature, performing classification and recognition based on the training semantic feature to obtain an initial training category, splicing the training semantic feature and the training global feature to obtain a spliced training feature, performing autocorrelation feature calculation based on the spliced training feature to obtain an autocorrelation feature, performing self-attention weight calculation based on the autocorrelation feature to obtain a self-attention weight, and adjusting the spliced training feature through the self-attention weight to obtain a fused training feature; and updating the initial feature fusion model based on the training global feature, the training semantic feature, the fused training feature, the initial training category, and a training sample category label, and performing a loop iteration to obtain a target fusion model.