Reproduce “Hi-Mamba: Hierarchical Mamba for Efficient Image Super-Resolution” on OpenResearch

About this paper

Junbo Qiao, Jincheng Liao, Wei Li, Yulun Zhang, Yong Guo, Yi Wen, Zhangxizi Qiu, Jiao Xie and 2 more

An efficient hierarchical Mamba model, Hi-Mamba, is introduced for image super-resolution, overcoming the computational overhead of multi-directional scanning in existing vision Mamba architectures. It achieves this by employing single-direction scanning within hierarchical blocks that capture multi-scale context, and uses a direction alternation strategy across groups of blocks to enrich spatial understanding. The approach yields improved reconstruction quality, for instance, a 0.21 dB PSNR gain on Urban100, while significantly reducing FLOPs compared to prior Mamba-based methods.

Key insights

• Direct application of multi-direction scanning in existing vision Mamba models significantly erodes the linear complexity advantage of SSMs for 2D image processing.
• Effective 2D spatial context modeling can be achieved efficiently by combining hierarchical feature extraction (local and region SSMs) with single-direction scanning.
• Alternating single-direction scans across cascaded blocks (DA-HMG) allows the network to aggregate diverse directional information and enrich spatial relationships without increasing parameters or per-block computational cost.

Summary by alphaXiv.