HyperTaFOR: Task-Adaptive Few-Shot Open-Set Recognition With Spatial-Spectral Selective Transformer for Hyperspectral...

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Open-set recognition (OSR) in hyperspectral images (HSI) faces challenges with limited training data and manual thresholding.
• Existing reconstruction-based and distance-based OSR methods struggle with efficient spectral-spatial information utilization.
• Accurate classification of known categories while rejecting unknown samples is crucial for HSI analysis.

Purpose of the Study:

• To develop a few-shot OSR framework for HSI that addresses data limitations and improves spectral-spatial feature extraction.
• To introduce a novel spatial-spectral selective transformer (S3Former) for enhanced feature learning.
• To enable task-adaptive rejection of unknown samples through meta-learning and negative prototype generation.

Main Methods:

• A few-shot OSR framework named HyperTaFOR is proposed for HSI.
• A spatial-spectral selective transformer (S3Former) is utilized for extracting spectral-spatial features.
• A meta-learning strategy with a negative prototype generation module (NPGM) is employed for adaptive sample rejection.

Main Results:

• HyperTaFOR demonstrates competitive classification and detection performance in open-set HSI scenarios.
• The S3Former effectively optimizes central pixel information while minimizing irrelevant spatial data.
• The NPGM generates task-adaptive rejection scores for flexible categorization of known classes and anomalies.

Conclusions:

• The proposed HyperTaFOR framework offers an effective solution for few-shot OSR in HSI.
• The S3Former architecture enhances the utilization of spectral-spatial information in OSR tasks.
• The meta-learning approach provides robust and adaptive rejection capabilities for unknown samples in HSI.