Reproduce “U-shaped Vision Mamba for Single Image Dehazing” on OpenResearch

About this paper

Zhuoran Zheng, Chen Wu

This paper introduces U-shaped Vision Mamba (UVM-Net), a deep learning model for single image dehazing that leverages the efficient Mamba architecture within a U-Net framework. UVM-Net achieves competitive or superior dehazing performance compared to Transformer-based models while significantly reducing computational overhead and demonstrating strong generalizability across multiple image restoration tasks.

Key insights

• The Mamba (SSM) architecture effectively captures long-range dependencies in image data with linear complexity, offering a more efficient alternative to Transformer's quadratic self-attention.
• Integrating Mamba within a U-Net framework via the Bi-SSM module, which processes features across both spatial and channel dimensions, optimizes its performance for dense prediction tasks like image restoration.
• The UVM-Net architecture demonstrates strong generalizability, performing well across diverse image restoration tasks including dehazing, low-light enhancement, and deraining, suggesting its potential as a versatile backbone.

Summary by alphaXiv.