Plkdcn: lightweight single-image super-resolution via partial large-kernel depthwise convolution

doi:10.1007/s00371-026-04742-1

Plkdcn: lightweight single-image super-resolution via partial large-kernel depthwise convolution

2026-09-23

0

PRE

AI

DOI：10.1007/s00371-026-04742-1

Original

Original request for help

Save

Abstract

En 中文

Deep convolutional neural networks (CNNs) have demonstrated exceptional performance in single-image super-resolution (SISR) tasks due to their superior ability to capture local features. However, achieving satisfactory performance generally requires sufficient network depth and width, and performance tends to degrade when the network is overly simplistic. Although transformers capable of modeling long-range dependencies are also commonly employed in super-resolution tasks, they incur significantly higher computational costs. Consequently, balancing network complexity and performance merits careful consideration. To address these challenges, we propose a partial large-kernel depthwise convolutional network (PLKDCN) for lightweight image super-resolution. This approach integrates multiple lightweight techniques within the feature modulation attention block (FMAB), which contains three main modules: PLKDC, DCFN and EA. Specifically, we incorporate the notion of “partial” processing into the modules and employ depthwise convolution to reduce the computational load of the convolutional layers. The partial large-kernel depthwise convolution (PLKDC) module leverages partial convolutions to remove redundant information from feature maps by applying convolutions only to a subset of the input channels, while compensating for potential performance degradation induced by lightweighting with large-kernel depthwise convolutions. In addition, the depthwise convolutional feedforward network (DCFN) is designed to capture long-range dependencies, and element-wise attention (EA) is introduced to enhance high-frequency information, further boosting performance. Extensive experiments demonstrate that the proposed PLKDCN achieves a favorable accuracy-complexity trade-off against state-of-the-art lightweight SISR methods. For $$\times $$ 4 super-resolution, the base PLKDCN model has only 415K parameters and 28 G FLOPs, with 16.5ms inference latency on an NVIDIA A40 GPU. Evaluations are conducted on five standard synthetic benchmarks and the real-world RealSR dataset, validating the effectiveness of the proposed design. The code and pre-trained models are available at https://github.com/sxdyyds/PLKDCN .

Keywords:

Single-image super-resolution

Lightweight image restoration

Efficient neural networks

Large-kernel depthwise convolution

Partial convolution

Frequency-domain loss

Resource-constrained deployment

Plkdcn: lightweight single-image super-resolution via partial large-kernel depthwise convolution

Abstract

Journal

Organization

Cited Papers