Test-Time Training for Hyperspectral Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Hyperspectral image (HSI) super-resolution (SR) lags behind RGB image SR due to challenges in modeling spectral band interactions and limited training data.
• Accurately modeling the complex spectral band interactions in HSIs is computationally intensive and difficult.
• Small and scarce datasets further hinder the development of effective HSI SR models.

Purpose of the Study:

• To address the limitations in Hyperspectral image (HSI) super-resolution (SR) research.
• To propose a novel test-time training method that improves HSI SR performance without complex spectral modeling.
• To enhance the diversity and quality of training data for HSI SR tasks.

Main Methods:

• Developed a self-training framework for test-time training, generating accurate pseudo-labels and LR-HR relationships.
• Proposed a new network architecture that learns HSI SR without explicit spectral band interaction modeling.
• Introduced a data augmentation technique called Spectral Mixup to increase training data diversity at test time.
• Collected a new, diverse HSI dataset covering various object categories.

Main Results:

• The proposed test-time training method significantly improved the performance of pre-trained HSI SR models.
• The novel network architecture and Spectral Mixup effectively addressed spectral interaction and data scarcity issues.
• The method demonstrated superior performance compared to existing HSI SR techniques across multiple datasets.

Conclusions:

• The developed test-time training framework offers a significant advancement in Hyperspectral image (HSI) super-resolution (SR).
• The approach effectively overcomes challenges related to spectral band complexity and limited data availability.
• This work provides a robust and efficient solution for improving HSI SR, paving the way for broader applications.