Unsupervised Structured-Noise Estimation and Real-Noise Validation

Develop unsupervised structured-noise estimation for AXS-Net and validate the resulting hyperspectral-image denoising approach on real-noise hyperspectral images, extending beyond synthetic structured-noise supervision and synthetically corrupted data.

Background

AXS-Net is trained with synthetic structured-noise supervision and is evaluated on synthetically corrupted hyperspectral images. The method explicitly estimates structured noise through its SBlock, but the reported experiments do not establish whether this decomposition can be learned without ground-truth structured-noise components.

The conclusion identifies two unresolved extensions: removing the reliance on synthetic supervision for structured-noise estimation and testing the method on real-noise hyperspectral imagery. These issues are important for determining whether the interpretable structured-noise branch remains effective under practical acquisition conditions and noise distributions that differ from the synthetic protocol.

References

Its main limitations are reliance on synthetic $S$ supervision, evaluation confined to synthetically corrupted data, and weaker MPSNR transfer on strongly shifted CAVE stripe and mixture cases; unsupervised structured-noise estimation and validation on real-noise HSIs are left for future work.

— AXS-Net: Interpretable Deep Unfolding for Hyperspectral Image Denoising via Spectral Basis Unmixing and Structured Noise Refinement  (2609.08777 - Guan et al., 8 Sep 2026) in Section Conclusion