基于混合感知与频率自适应门控网络的轻量图像超分辨率重建算法

Lightweight Image Super-Resolution Reconstruction Algorithm Based on Hybrid Perception and Frequency-Adaptive Gating Network

摘要: 基于Transformer的方法凭借其卓越的长距离依赖建模能力，在单图像超分辨率领域取得了显著进展。然而，现有的轻量级Transformer架构在追求计算效率时，往往通过通道压缩或稀疏窗口机制来降低计算负担，此类策略削弱了局部特征的空间连续性，且自注意力机制固有的低通滤波特性限制了网络对高频纹理细节的恢复能力。为了解决上述频率偏差与局部信息丢失的问题，本文提出了一种基于混合感知与频率自适应门控网络的轻量图像超分辨率重建算法HPG-SR。首先，本文设计了混合感知门控注意力模块，通过并行使用局部感知分支和可学习的门控机制，在保留大窗口全局感受野的同时，显式地强化局部高频细节。其次，本文提出了多尺度门控前馈网络，利用双路多尺度卷积和上下文门控替代传统的静态激活函数，增强了网络对不同频率特征的自适应选择能力。最后，提出了对比度感知特征细化模块，利用标准差统计量强化对纹理丰富区域的特征响应。在五个基准数据集上的广泛实验表明，HPG-SR在参数量和计算量相当的情况下，性能优于当前最先进的轻量级SR方法。特别是在纹理复杂的Urban100数据集上，该算法展现出了更佳的细节恢复能力。

Abstract: Transformer-based methods have achieved significant progress in single image super-resolution due to their superior ability to model long-range dependencies. However, existing lightweight Transformer architectures often employ channel compression or sparse window mechanisms to reduce computational burden, which inevitably weakens the spatial continuity of local features. Furthermore, the inherent low-pass filtering nature of the self-attention mechanism limits the network’s capacity to recover high-frequency texture details. To address the issues of frequency bias and local information loss, this paper proposes a lightweight image super-resolution reconstruction algorithm based on a Hybrid Perception and Frequency-Adaptive Gating network, named HPG-SR. First, a Hybrid Perception Gated Attention module is designed. By utilizing a parallel local perception branch and a learnable gating mechanism, it explicitly enhances local high-frequency details while retaining the global receptive field of large windows. Second, a Multi-Scale Gated Feed-Forward Network is proposed, which employs dual-path multi-scale convolutions and context gating to replace traditional static activation functions, thereby enhancing the network’s adaptive selection capability for features across different frequencies. Finally, a Contrast-Aware Feature Refinement module is introduced to strengthen feature responses in texture-rich regions using standard deviation statistics. Extensive experiments on five benchmark datasets demonstrate that HPG-SR outperforms state-of-the-art lightweight SR methods with comparable parameters and computational complexity. Particularly on the texture-complex Urban100 dataset, the proposed algorithm exhibits superior detail recovery capability.