一种基于混合注意力机制的图像特征提取方法
By designing an image feature extraction method with a hybrid attention mechanism, and utilizing dynamic clustering and K-means clustering updates, the problem of the lack of global information interaction in the Transformer model is solved, thereby improving the image feature extraction capability and the performance of dense prediction tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-08-18
- Publication Date
- 2026-07-17
AI Technical Summary
The Transformer model based on local attention mechanism lacks explicit global information interaction, resulting in insufficient ability to extract global image features, which limits its application in dense prediction tasks.
Design an image feature extraction method based on a hybrid attention mechanism. The key value matrix is divided into cluster centers through a dynamic clustering mechanism, local and global attention matrices are constructed, and global information interaction with linear complexity is achieved through a hybrid attention module. Combined with K-means clustering to update the cluster assignment, a hybrid attention output is generated.
It improves the performance of the Transformer model in dense prediction tasks, especially significantly improving segmentation accuracy and reducing computational complexity in semantic segmentation tasks.
Smart Images
Figure CN121073906B_ABST