DBFF-SRNet: an image super-resolution method based on dual-branch encoding and frequency-domain fusion

With the increasing demand for high-resolution images in remote sensing, medical imaging, and computer vision, image super-resolution has become a crucial technique for enhancing image quality. However, existing methods often face a trade-off between detail restoration and structure preservation and struggle to effectively fuse multimodal information, limiting further improvements. To address these challenges, this paper proposes DBFF-SRNet, a dual-branch frequency fusion network that separately extracts structural and texture features through dedicated encoders, and integrates them using a spatial-spectral joint fusion module along with a multi-level wavelet pyramid frequency enhancement module, enabling efficient multimodal feature fusion and detail recovery. Experimental results demonstrate that DBFF-SRNet outperforms state-of-the-art methods on benchmark datasets including Set5, Set14, BSD100, and Urban100. Specifically, on Set5, DBFF-SRNet achieves a PSNR of 32.55 dB, improving by 0.35 dB over the second-best method, with an SSIM of 0.8927 and a reduced LPIPS of 0.0574, significantly enhancing the visual quality and structural coherence of reconstructed images. The proposed model not only provides an effective solution for high-quality multimodal image reconstruction but also opens new avenues for integrating frequency-domain and spatial-domain deep learning approaches in super-resolution research.