Abstract: The main challenge of single image super-resolution (SISR) tasks is how to recover the high-frequency details of low-resolution images while reducing computational complexity and avoiding artifacts. In order to effectively solve these problems, the multi-scale large-kernel network WEMA-Net fused with wavelet convolution is proposed. Firstly, the network combines wavelet transform convolution with large kernel decomposition strategy, and performs convolution operations in the frequency domain, which expands the receptive field and reduces the computational complexity. Then, a multi-scale grouping mechanism is introduced to capture long-distance dependence and local texture information at different scales, and the wavelet convolution is combined to reduce the artifacts that may be caused by large kernel convolution. Then, an adaptive gated spatial attention unit is proposed, which can dynamically adjust the attention weight according to the features of different regions, so as to optimize the fusion of local and global features, and enhance the recovery ability of details and edges. In addition, the modulation convolution unit module is introduced, which further improves the flexibility of feature extraction by improving the robustness to outliers. Tested on datasets such as Set5, Set4, B100, Urban100 and Manga109, the experimental results show that WEMA-Net has an improvement in peak signal-to-noise ratio (PSNR) compared with lightweight models such as RFDN, BSRN, LKDN, RepRFN, OSFFNet, etc. In the 4x super-resolution, Urban100 and Manga109 are 0.18 dB and 0.2 dB, respectively, compared to LKDN. Experimental results verify the superiority of the network in image detail recovery, computational efficiency and robustness, indicating that the proposed method has a wide application prospect.

摘要： 单图像超分辨率（SISR）任务当前面临的主要难题在于，如何在恢复低分辨率图像的高频细节同时降低计算复杂度和避免伪影的产生。为了有效解决这些问题，提出了融合小波卷积的多尺度大核网络WEMA-Net。该网络首先把小波变换卷积与大核分解策略结合，在频域内进行卷积操作，扩展了感受野并降低了计算复杂度。再设计多尺度分组机制，在不同尺度上捕获长距离依赖和局部纹理信息，结合小波卷积减少大核卷积可能导致的伪影。接着提出了自适应门控空间注意力单元，能够根据不同区域的特征动态调整注意力权重，从而优化局部和全局特征的融合，增强了细节和边缘的恢复能力。另外还设计了调制卷积单元模块，通过提升对异常值的鲁棒性，进一步提高了特征提取的灵活性。在Set5，Set4，B100，Urban100和Manga109等数据集上测试，实验结果表明，WEMA-Net在峰值信噪比（PSNR）上相较于轻量化模型RFDN，BSRN，LKDN，RepRFN，OSFFNet等上均有提升。在4倍超分中，相比于LKDN的Urban100和Manga109分别提高了0.18dB和0.2dB。实验结果验证了该网络在图像细节恢复、计算效率及鲁棒性方面的优越性，表明该方法具有广泛的应用前景。