Small target detection method based on Manhattan attention and multi-scale feature query
By introducing the Manhattan attention and decay matrix into small object detection, limiting the attention range of Query, the problem of low accuracy and recall in complex environments is solved, and higher detection accuracy is achieved.
CN119693622APending Publication Date: 2025-03-25CHONGQING UNIV OF POSTS & TELECOMM
0 Cites 1 Cited by
Patent Information
- Application Number
- CN202411839726.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-13
- Publication Date
- 2025-03-25
AI Technical Summary
Technical Problem
In complex environments, the accuracy and recall rate of small target detection are low, mainly because Query's attention range is too wide, resulting in the small target information being masked by the information of significant targets.
Method used
The small object detection method based on Manhattan attention is adopted. By introducing a decay matrix into Query, the attention range of Query is limited, so that it only focuses on tokens containing small object information in a small range, thereby avoiding information masking.
Benefits of technology
The detection accuracy and recall of small targets are significantly improved, ensuring that the model can effectively identify small targets in complex environments.
✦ Generated by Eureka AI based on patent content.
Abstract
The invention provides a small target detection method based on Manhattan attention and multi-scale feature query. A specific network of the small target detection method comprises a backbone network, a context sensing Query generation module, an encoder based on deformable self-attention and a decoder based on deformable and Manhattan attention. In a decoder based on deformable and Manhattan attention, deformable attention is used to perform global target detection on feature vectors output by an encoder. The multi-layer Manhattan attention is used to perform local target detection on the feature vector. The multi-layer Manhattan attention has the characteristic of multi-scale Manhattan attention. The small target detection method based on Manhattan attention and multi-scale feature query provided by the invention can effectively improve the accuracy and recall rate of small targets in a scene in which large, medium and small targets exist at the same time.
Need to check novelty before this filing date? Find Prior Art