Diff-Mamba: A diffusion-Mamba framework for hyperspectral image classification

doi:10.1016/j.neucom.2025.131930

Diff-Mamba: A diffusion-Mamba framework for hyperspectral image classification

2025-10-27

0

PRE

AI

DOI：10.1016/j.neucom.2025.131930

Original

Original request for help

Save

Abstract

En 中文

• We propose a novel framework that integrates diffusion models with Mamba to fully explore and utilize rich contextual semantic information, significantly improving classification performance. To the best of our knowledge, this is the first work to construct a diffusion network using Transformer blocks for hyperspectral image classification. • To model complex spectral-spatial distribution relationships, we design the spectral-spatial diffusion feature generation module. This module encompasses forward and reverse diffusion processes, utilizing a meticulously designed spectral-spatial diffusion network to extract spectral-spatial diffusion features that encapsulate distribution information. The diffusion features can enhance cross-sample perception by fitting real HSI data distribution from a generative perspective. • We propose the center pixel-driven semantic token generator that dynamically adjusts the weights of the target and other pixels in the patch, significantly enhancing generation and focusing of discriminative features.

Diff-Mamba: A diffusion-Mamba framework for hyperspectral image classification

Abstract

Journal

Organization

Cited Papers