MAANet: A lightweight multi-axis adaptation network for efficient image super-resolution

Prevailing lightweight SR architectures typically stack homogeneous blocks, limiting feature diversity and multi-dimensional cue capture. Furthermore, these approaches face a spatial aggregation dilemma: local window mechanisms are efficient but content-agnostic, whereas clustering-based approaches are flexible yet computationally expensive. To address these issues, we propose the Multi-Axis Adaptation Network (MAANet). Built upon the Multi-Axis Processing Block (MAPB), our method adopts a “global-local” multi-level hierarchical structure to integrate features from three distinct perceptual dimensions, termed “Axes”. (1) Globally, the Multi-Axis Core Module (MACM) captures long-range context in parallel using the Dynamic Kernel Synthesis Block (DKSB) to generate content-adaptive kernels on the Kernel Axis, and the Spectral-Spatial Modulation Unit (S2MU) to model global periodic priors on the Frequency Axis. (2) Locally, the Adaptive Tiling Transformer Block (ATTB) on the Spatial Axis efficiently allocates computational resources to texture-rich areas for precise refinement. (3) We introduce a Dual-Dynamic Fusion Upsampler Block (D2FUB) to extend the content-aware reconstruction capability to the final upsampling stage, surpassing the widely used PixelShuffle operator in reconstruction quality with lower complexity. Extensive experiments demonstrate that MAANet achieves state-of-the-art results in both reconstruction performance and computational efficiency. Notably, compared to SRConvNet-L (×4), MAANet improves PSNR by up to 0.28 dB with only about 70% of FLOPs and 59% of parameters, while running 4.7 times faster than CATANet (×4). Code is available at: https://github.com/chrimy666999/MAANet.