RMRDN: Recurrent multi-receptive residual dense network for image super-resolution

Reconstructing fine textures and structures from low-resolution images remains a central challenge in super-resolution (SR). Existing CNN-based SR models often suffer from limited receptive fields, weak long-range dependency modeling, and insufficient use of hierarchical features. To address these limitations, we propose a Recurrent Multi-Receptive Residual Dense Network (RMRDN) comprising three novel modules: (1) a Recurrent Multi-Receptive Residual Dense Block (RMRDB) for capturing rich contextual information; (2) a Residual Dense LSTM (RDLSTM) for long-range dependency modeling; and (3) a Relevant Feature Booster Block (RFBB) for effective hierarchical feature utilization. Extensive experiments on five benchmark datasets demonstrate that RMRDN outperforms existing methods by producing sharper textures and more accurate structural details. For ×4 upscaling, our proposed model outperforms the second-best SR method, achieving gains of +0.10 dB on Set5, +0.11 dB on Set14, +0.13 dB on BSD100, +0.06 dB on Urban100, and +0.11 dB on Manga109, respectively.