Fundus Image Segmentation With Dynamic Weighted Attention Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fundus image segmentation methods using deep learning are complex, time-consuming, and resource-intensive, often requiring deep network models that consume excessive GPU resources while lacking robustness and intuitive lesion display.
Innovation Solution
A method and system utilizing a dynamic weighted attention mechanism for hierarchical feature fusion on a shallow network model, reducing algorithm complexity and resource usage while maintaining recognition accuracy by performing downsampling, weighted attention fusion, and classification convolution operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep network models are used to extract richer features, then recognition accuracy is improved, but algorithm complexity and running time increase
Solution Approach 1:
The patent applies dynamic weighted attention mechanisms that adaptively adjust feature weights during processing. The attention weights are dynamically computed based on feature importance, allowing the model to focus on relevant features without increasing network depth. This resolves the contradiction by providing adaptive feature selection in a shallow architecture.
Solution Approach 2:
The patent changes the parameter representation by introducing attention weight parameters that control feature fusion. Instead of increasing network depth, the model uses learnable attention parameters to enhance feature extraction capability, maintaining shallow architecture while improving recognition accuracy through parameter optimization.
2Measurement precision
If deep network models are used to extract richer features, then recognition accuracy is improved, but running time increases
Solution Approach 1:
The dynamic attention mechanism computes weights adaptively during feature fusion, allowing the model to quickly identify and prioritize important features without processing through multiple deep layers. This reduces computational steps and running time while maintaining accurate feature extraction.
Solution Approach 2:
The patent extracts only the most important features using attention mechanisms rather than processing all features through deep networks. By selectively extracting and fusing key features from different layers, the model achieves accurate recognition with fewer computational operations, reducing running time.
3Measurement precision
If deep network models are used to extract richer features, then recognition accuracy is improved, but GPU resource consumption increases
Solution Approach 1:
The patent optimizes GPU resource usage by changing how features are processed - using attention weight parameters to control feature fusion instead of increasing network depth. This parameter-based approach reduces the computational burden on GPU while maintaining feature extraction effectiveness and recognition accuracy.
Solution Approach 2:
The patent uses feature copying and fusion from different layers through attention mechanisms rather than creating entirely new features through deep processing. By copying and selectively fusing features from shallow and deep layers, the model achieves rich feature representation with reduced GPU computation requirements.
4Productivity
If shallow network models are used, then algorithm complexity and running time are reduced, but feature extraction capability deteriorates
Solution Approach 1:
The dynamic attention mechanism compensates for the shallow architecture by adaptively weighting and fusing features from different layers. This dynamic feature selection enhances the effective feature extraction capability of the shallow model, matching or exceeding deep models while maintaining algorithm efficiency.
Solution Approach 2:
The patent merges features from multiple layers through attention-based fusion, combining the strengths of shallow and deep features in a unified representation. This merging process enhances feature extraction capability without requiring increased network depth, maintaining algorithm efficiency while improving feature quality.
Data Source
AI summary
The present disclosure provides a method and system for detecting a fundus image based on a dynamic weighted attention mechanism. Lesion information in a fundus image of a premature infant is detected using a fundus image segmentation model. First, the fundus image is consecutively downsampled. Dynamical weighted attention fusion is performed on an obtained downsampling feature and an obtained downsampling feature of an adjacent layer. The weighted and fused features are fused with an output feature of a corresponding upsampling layer. Finally, a classification convolution operation is performed on an output of an n-th upsampling layer to obtain a lesion probability for each pixel. The present disclosure performs hierarchical feature fusion on a shallow network model using the dynamic weighted attention mechanism, which can reduce complexity of algorithm design, shorten a running time of an algorithm, and reduce excessive occupation of graphics processing unit (GPU) resources while ensuring recognition accuracy.