Work

Efficient Hierarchical Non-Local Retrieval for Arbitrary-Scale Image Super-Resolution

Year: 2026
Type: article
Abstract: Implicit image functions enable a single model to perform arbitrary-scale image super-resolution. However, most existing methods rely mainly on local features near each query coordinate, limiting their ability to exploit repeated textures and long-range structural information. Dense non-local matching can complement local features but requires expensive exhaustive feature comparisons on high-resolution feature maps. To address this problem, we propose Sparse Attention with Coarse-to-Fine Aggregation (SACA), an efficient hierarchical non-local retrieval method for arbitrary-scale image super-resolution. SACA performs global similarity search only at the coarsest feature level to identify potentially relevant regions, and progressively propagates and re-evaluates the selected candidates at finer levels. A local refinement module further adjusts the sampling locations around the retrieved candidates before feature aggregation. Experiments on DIV2K, Urban100, and B100 demonstrate competitive reconstruction performance at both in-scale and out-of-scale factors. On Urban100, SACA achieves consistent improvements over LTE, with a peak signal-to-noise ratio (PSNR) gain of 0.1693 dB at ×8. Meanwhile, the proposed coarse-to-fine retrieval reduces the computational cost by up to 82.4% compared with exhaustive fine-level matching over the evaluated input sizes. These results demonstrate that SACA provides an effective balance between reconstruction quality and computational efficiency.
Language: English
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7234612726
FWCI:
Cites: 31
Cited by:
Related to: 10