Prototype perception sample adaptive adjustment method and system for long-tail distribution out-of-distribution image detection
By introducing a prototype-aware ID feature calibration module and an OOD logits adjustment module, the technical problems in long-tailed distribution out-of-image detection are solved through prototype-aware techniques, achieving better detection results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2026-03-25
- Publication Date
- 2026-07-17
AI Technical Summary
Existing methods lack fine-grained learning based on sample-level differences in image detection outside of long-tailed distributions, resulting in an imbalance in the model's learning of different samples within each category, which affects the model's generalization performance.
The prototype-aware ID feature calibration module (PA-IFC) is used to model the diversity of intra-class features, and the OOD logits adjustment module (PA-OLA) is used to adaptively adjust the model to learn OOD samples, thereby improving the accuracy of ID image classification and the effect of OOD detection.
By calibrating the feature representation of each sample class using the PA-IFC module, the decision boundary is broadened. By reshaping the logits distribution of the model on OOD data using the PA-OLA module, the accuracy of image detection outside the long-tail distribution and the effect of OOD detection are improved.
Smart Images
Figure CN122416092A_ABST