TL;DR
A lightweight dual-dimension modulation aggregation network, which combines channel-wise and spatial feature interactions to achieve more accurate reconstruction, and shows that DMANet achieves competitive reconstruction performance with lower model complexity and runtime overhead.
Significant progress has been made in remote sensing image super-resolution based on deep neural networks. However, existing methods typically suffer from parameter redundancy and high computational costs, making them difficult to deploy on resource-constrained edge devices. Moreover, the image reconstruction process often involves phenomena such as texture blurring and edge distortion, which degrade the restoration quality. To address these issues, we propose a lightweight dual-domain attention aggregation network (LDANet), aiming to achieve image super-resolution with both high-efficiency and high-quality. Specifically, LDANet is composed of multiple cascaded dual domain attention-guided feature aggregation blocks (DAGBs). Within each DAGB, we propose the variance-aware spatial attention module, which enhances the extraction capability of nonlocal salient features through a dual-path mechanism that integrates parallel adaptive pooling and feature variance modulation strategies, thereby improving the discriminability of spatial features while effectively suppressing noise interference. To further optimize pixel-level detail expression, we propose the pixel-embedding channel attention module, which achieves cross-channel global context awareness by jointly modeling pixel-level spatial relationships and channel-wise self-attention, thereby enhancing texture fidelity and edge clarity. Moreover, asymmetric convolutional gated feed-forward network is introduced to enhance the feature aggregation and local context mixing capabilities within the DAGB. Comparative experiments demonstrate that LDANet achieves superior reconstruction performance on multiple datasets with smaller model sizes, exhibiting better potential for practical applications. The source code of LDANet will be released at https://github.com/AHUT-MILAGroup/LDANet.
Transformer-based super-resolution methods achieve competitive performance by capturing non-local information through self-attention. Nevertheless, the computation of the self-attention introduces heavy computational overhead, and its inherent low-pass characteristic restricts the learning of local details. To address these problems, we propose an effective dual-branch multi-perspective modulation network (DMMN) for efficient infrared image super-resolution. Specifically, we design a multi-scale feature modulation enhancement unit (MFMEU) to capture cross-scale spatial features and a frequency-domain cross-correlation patch modulation unit (FCPMU) to explore global feature representations. We further develop an efficient bidirectional cross modulation unit (BCMU) to promote feature interaction between outputs of MFMEU and FCPMU. Extensive experimental results verify that DMMN achieves a competitive trade-off between reconstruction accuracy and computational efficiency. For instance, compared with the ×4 SRFormer-light, the proposed DMMN obtains an average gain of 0.08 dB in PSNR across five public test datasets, runs 2.7× faster, and uses only about 24% of the FLOPs.
An efficient aligned kernel network (AKNet) is proposed, which innovatively employs super-large convolution kernels to capture global receptive fields with minimal computational overhead, effectively mimicking the long-range dependency modeling of transformers.
Image super-resolution (ISR) has emerged as a critical computer vision task aimed at reconstructing high-resolution visual information from low-resolution inputs. Although deep learning-based approaches have significantly improved reconstruction quality, many existing architectures suffer from high computational complexity, excessive parameter requirements, and limited efficiency in real-time deployment scenarios. This research presents a Lightweight Transformer-Fourier Fusion Framework for Efficient Image Super-Resolution, designed to integrate the long-range dependency modeling capability of transformers with the frequency-domain representation advantages of Fourier-based feature processing. The proposed framework is theoretically positioned around efficient feature extraction, adaptive attention learning, and frequency-aware reconstruction. Transformer-based modules enhance spatial relationship modeling, while Fourier convolution mechanisms improve the preservation of high-frequency image details with reduced computational overhead. The methodology combines lightweight residual feature refinement, transformer-driven contextual enhancement, and frequency-domain fusion to achieve an optimized balance between reconstruction accuracy and computational efficiency. The study analyzes the limitations of conventional convolutional, generative, and attention-based super-resolution approaches and establishes the importance of hybrid architectures for next-generation ISR systems. The proposed framework provides a scalable solution for applications requiring efficient image enhancement, including mobile imaging, medical visualization, remote sensing, and intelligent surveillance systems.
Single-image super-resolution (SISR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) observations while preserving structural information and high-frequency detail. Although the Hybrid Multi-Axis Network (HMA) effectively combines local and nonlocal attention, its shallow input representation still mixes low-frequency structure, directional detail, and noise-like high-frequency components. This study investigates whether an explicit frequency prior can be introduced before the HMA backbone without substantially increasing computational cost. Two discrete-wavelet-transform front-ends are examined under the ×2 setting. HMA-WSB uses lightweight subband-specific processing and weighted fusion before a shared HMA backbone, whereas HMA-MSB introduces asymmetric multi-subband branches and cross-band fusion. Evaluation includes the reported external 15-image experiment, selected-image pilots from Set5, Set14, BSD100, and Urban100, and supplementary medical and texture-domain samples. The results show small, content-dependent differences rather than a consistent reconstruction advantage: the proposed variants are slightly favorable on several images containing dense multidirectional detail, but the original HMA remains stronger on other natural, medical, and periodic-texture samples. Computational analysis on an NVIDIA GeForce RTX 5070 with a 64×64 low-resolution input shows that HMA-WSB increases measured inference latency by 1.637% with negligible parameter and memory overhead. HMA-MSB increases latency by 5.078%, parameter count by 1.463%, and estimated FLOPs by 0.489%. These findings indicate that wavelet-guided subband processing is compatible with HMA and that WSB provides the more computationally economical extension. However, because the standard-dataset evaluation is based on selected images and a complete component-level ablation is not available, the results should be interpreted as preliminary evidence of a content-dependent quality-cost trade-off rather than proof of broad superiority.
Sparse attention mechanisms, which score all token pairs but propagate only the strongest, now underpin the most efficient Transformers for lightweight image super-resolution. This paper observes that sparsification changes what it means to improve such a network. A dense attention layer has one place where representation quality matters: the aggregation of attended features. A sparse layer has two, because the top-k operator first decides which tokens survive and only then decides what to do with them, and a token discarded at the selection stage cannot be recovered downstream. Selection quality and aggregation quality are therefore separable targets, addressed by modules placed before and after the attention respectively. We test this by pairing a dual-branch spatial enhancement on the input of a progressive focused attention with a wavelet-domain modulation on its output, forming SFMformer. Measuring each module alone and jointly over all fifteen benchmark-scale pairs, we find their gains are not additive: the joint gain exceeds the sum of the individual gains on nine pairs, and the sign of the discrepancy is predicted by how much the weaker module contributes on its own (r = -0.72), so the two compound when they relieve different constraints and overlap when they relieve the same one. Enabling spectral modulation once per block rather than once per layer retains the effect at roughly one-sixth of its cost, keeping the model below one million parameters at every scale. SFMformer ranks first on 28 of 30 PSNR/SSIM entries across five benchmarks and three upscaling factors. We report the cases where the pairing does not help, and deploy the model on a Raspberry Pi 5 to confirm the design is practical under tight resource budgets.
We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.
