Hierarchical representation allocation for efficient image super-resolution

Single image super-resolution (SR) aims to reconstruct a high-resolution image from its low-resolution counterpart. Although recent deep learning-based SR methods have achieved remarkable progress, obtaining favorable reconstruction quality under strict computational budgets remains challenging. In particular, lightweight SR models often process heterogeneous spatial cues, such as large-scale structural patterns and local high-frequency details, within a largely unified enhancement stream, which may lead to suboptimal allocation of limited representational capacity. To address this issue, we propose an efficient SR network based on hierarchical representation allocation. Specifically, we design an Allocation-Guided Spatial Modulation (AGSM) module that explicitly performs branch-level pre-allocation before branch-specific processing. Within AGSM, Structure-Oriented Multi-Scale Organization (SOMSO) enhances structural responses across projected scales, while Detail-Oriented Statistical Modulation (DOSM) captures localized detail-sensitive responses through lightweight statistical modulation. Furthermore, we develop an Allocation-Conditioned Fine-Grained Local Refinement (ACFGLR) module, which preserves continuous local quadrants as the local refinement basis and calibrates region-wise enhancement using channel response, spatial response, local variance, and inherited allocation prior. In this way, the proposed method consistently organizes enhancement from branch level to region level under lightweight constraints. Experimental results on several benchmark datasets demonstrate that the proposed method achieves a favorable trade-off between reconstruction performance and computational efficiency, outperforming or remaining competitive with state-of-the-art lightweight SR methods while maintaining low complexity.