从序列到视觉：状态空间模型与视觉架构融合的全面综述与前沿展望

From Sequence to Vision: A Comprehensive Review and Frontier Prospects of the Integration of State Space Models and Vision Architectures

摘要: 状态空间模型(State Space Models, SSMs)作为控制理论和系统识别领域的经典方法，近年来在深度学习的推动下焕发出新的生命力。特别是2023年提出的Mamba架构，通过引入输入依赖的选择性机制和高效的并行扫描算法，在自然语言处理领域取得了突破性进展，挑战了Transformer的主导地位。这一成功自然地引发了计算机视觉领域的研究兴趣：能否将这种高效的序列建模范式应用于具有空间结构的视觉数据？本综述系统性地回顾了SSMs在计算机视觉领域的迁移、融合与应用进展。我们首先深入剖析了结构化SSMs (特别是Mamba)的核心原理与计算特性；全面梳理了SSMs与视觉架构融合的三种主要范式：纯SSM骨干网络、混合架构以及多模态融合模型；最后，我们探讨了这一领域当前面临的开放性问题，并展望了未来可能的研究方向。本文旨在为研究者提供一个关于SSM在视觉任务中应用的全面视角，推动更高效、更强大的视觉基础模型的发展。

Abstract: State Space Models (SSMs), a classical methodology in control theory and system identification, have been revitalized by recent advances in deep learning. In particular, the Mamba architecture introduces input-dependent selective mechanisms and efficient parallel scan algorithms, achieving competitive long-sequence modeling performance while maintaining linear complexity. This review systematically surveys the migration, integration, and application of SSMs in computer vision. It analyzes the principles and computational characteristics of structured SSMs, especially Mamba; summarizes three major paradigms for integrating SSMs with vision architectures, including pure SSM backbones, hybrid architectures, and multimodal fusion models; and discusses open challenges and future research directions. The paper aims to provide researchers with a comprehensive perspective on applying SSMs to visual tasks and to promote the development of more efficient and powerful vision foundation models.