Attention based deep feature fusion for hyperspectral target detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral target detection faces challenges like high dimensionality, background variability, and class imbalance.
• Existing methods struggle with effectively integrating global context and local details.

Purpose of the Study:

• To propose a novel framework, Self and Cross Attention Feature Fusion (SCAFF), for robust hyperspectral target detection.
• To address challenges of spectral dimensionality, background variability, and class imbalance.

Main Methods:

• A patch-based framework integrating an attention-driven branch (for intra- and inter-patch dependencies) and a convolutional branch (for local patterns).
• Joint use of Self-Attention and Cross-Attention to focus on potential target regions and mitigate class imbalance.
• An enhanced guided filter for refining detection map spatial consistency in post-processing.

Main Results:

• SCAFF demonstrated superior accuracy and stability across four benchmark datasets.
• Achieved performance improvements of up to 15% compared to state-of-the-art methods.
• Validated the effectiveness and robustness of the proposed method for practical applications.

Conclusions:

• SCAFF effectively balances global contextual modeling with fine-grained local representation.
• The attention mechanisms adaptively guide focus, enhancing discriminability and mitigating class imbalance.
• The proposed method offers a robust solution for challenging hyperspectral target detection tasks.