HyperCASR: Spectral-Spatial Open-Set Recognition With Category-Aware Semantic Reconstruction for Hyperspectral...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Open-set recognition (OSR) in hyperspectral imagery (HSI) aims to classify known classes while rejecting unknown samples.
• Existing reconstruction-based methods struggle with noise and inter-class confusion in HSI.
• Leveraging spectral-spatial information for HSI OSR remains a significant challenge.

Purpose of the Study:

• To propose HyperCASR, an innovative framework for hyperspectral imagery open-set recognition.
• To enhance the extraction of spectral-spatial features and reduce noise and confusion.
• To accurately classify known classes and effectively reject unknown classes in HSI.

Main Methods:

• Developed a grouped spectral-spatial retentive transformer (GSSRT) for feature extraction.
• Integrated a grouped pixel embedding (GPE) and spatial retentive attention (SRA) mechanism within GSSRT.
• Employed a class-aware semantic reconstruction (CASR) module with independent autoencoders (AEs) for each known class.

Main Results:

• The proposed GSSRT enhances the extraction of spatial-spectral information.
• The CASR module effectively mitigates noise interference and inter-class confusion.
• HyperCASR demonstrated significantly improved classification performance for both known and unknown classes on benchmark datasets.

Conclusions:

• HyperCASR offers a robust solution for hyperspectral imagery open-set recognition.
• The framework successfully addresses limitations of existing reconstruction-based approaches.
• Experimental results validate the superiority of HyperCASR over state-of-the-art methods.