Wildlife image feature enhancement method and device based on semantic guidance
By generating semantic prototype vectors using a large language model and leveraging a cross-modal attention mechanism to enhance the visual features of wildlife images, this approach addresses the issues of poor feature stability and insufficient inter-class discriminability caused by pose and occlusion, achieving efficient and accurate feature enhancement for identification.
Patent Information
- Application Number
- CN202610529019.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-17
AI Technical Summary
Existing wildlife image identification feature enhancement techniques suffer from poor feature stability due to pose and occlusion, insufficient enhancement of subtle inter-class features, and limited generalization ability. Furthermore, existing semantic assistance schemes are weak in guidance and have low enhancement efficiency.
By mining and encoding standardized semantic knowledge of species through a large language model, semantic prototype vectors are generated. A cross-modal attention mechanism is used to target and enhance visual features, focusing on key identification regions and suppressing background noise. Furthermore, a visual-semantic fusion module is used to integrate complementary semantic information, thus constructing an end-to-end bimodal network framework and achieving a decoupled design between training and inference.
It improves the efficiency and accuracy of wildlife image identification, enhances the robustness of features and inter-class distinguishability, and reduces computational overhead.
Smart Images
Figure CN122416487A_ABST