Spatial-gate self-distillation network for efficient image super-resolution

Abstract

The balanced extraction of both non-local and local features represents a critical requirement for effective image super-resolution (SR). While transformer-based self-attention (SA) mechanisms demonstrate superior non-local modeling capabilities, their substantial computational demands limit practical deployment. To address this efficiency-performance trade-off, the Spatial-Gate Self-Distillation Network (SGSDN) implements a dual-capacity architecture combining: an SA-like (SAL) module employing strategically dilated 1D depthwise convolutions in horizontal and vertical orientations for efficient non-local feature extraction, and a lightweight local spatial-gate (LKG) block optimized for local detail preservation. Moreover, the proposed spatial-gate self-distillation block (SGSDB) further enhances performance through an optimized distillation structure that simultaneously processes both feature types while minimizing memory overhead. Experimental results demonstrate SGSDN’s superior performance-complexity balance, with benchmark evaluations showing comparable accuracy to SwinIR-light while requiring only 25% of the computational resources (FLOPs) and 25% of parameters, attributable to its avoidance of computationally intensive matrix operations.