Deep Back-Projection Networks (DBPN) are proposed for single-image super-resolution (SR). Unlike traditional feed-forward networks that learn a non-linear mapping from low-resolution (LR) to high-resolution (HR) images, DBPN uses iterative up- and down-sampling layers to provide error feedback, improving the reconstruction of HR images. The network includes mutually connected up- and down-sampling stages that represent different types of image degradation and HR components. By concatenating features across these stages (Dense DBPN), the network achieves superior performance, especially for large scaling factors like 8×. DBPN introduces an error feedback mechanism that calculates both up- and down-projection errors to guide the reconstruction process. It also uses dense connections between stages to encourage feature reuse, enhancing the network's ability to reconstruct HR images. The network is trained end-to-end, with each stage alternating between up- and down-sampling operations. This architecture allows the network to preserve HR components by learning various up- and down-sampling operators and generating deeper features. The proposed DBPN outperforms existing SR methods in terms of PSNR and SSIM on multiple datasets, including Set5, Set14, BSDS100, Urban100, and Manga109. It achieves higher performance than state-of-the-art methods like EDSR and LapSRN, particularly for large scaling factors. The network is also more efficient in terms of parameters, with D-DBPN having significantly fewer parameters than EDSR while maintaining comparable PSNR. The DBPN architecture is modular, allowing for easy adjustment of the number of stages and depth. It is effective in preserving HR components and generating high-quality images, even with shallow networks. The results show that DBPN is capable of reconstructing HR images with high fidelity, especially for large scaling factors. The network's ability to incorporate feedback and dense connections makes it a powerful approach for SR tasks.Deep Back-Projection Networks (DBPN) are proposed for single-image super-resolution (SR). Unlike traditional feed-forward networks that learn a non-linear mapping from low-resolution (LR) to high-resolution (HR) images, DBPN uses iterative up- and down-sampling layers to provide error feedback, improving the reconstruction of HR images. The network includes mutually connected up- and down-sampling stages that represent different types of image degradation and HR components. By concatenating features across these stages (Dense DBPN), the network achieves superior performance, especially for large scaling factors like 8×. DBPN introduces an error feedback mechanism that calculates both up- and down-projection errors to guide the reconstruction process. It also uses dense connections between stages to encourage feature reuse, enhancing the network's ability to reconstruct HR images. The network is trained end-to-end, with each stage alternating between up- and down-sampling operations. This architecture allows the network to preserve HR components by learning various up- and down-sampling operators and generating deeper features. The proposed DBPN outperforms existing SR methods in terms of PSNR and SSIM on multiple datasets, including Set5, Set14, BSDS100, Urban100, and Manga109. It achieves higher performance than state-of-the-art methods like EDSR and LapSRN, particularly for large scaling factors. The network is also more efficient in terms of parameters, with D-DBPN having significantly fewer parameters than EDSR while maintaining comparable PSNR. The DBPN architecture is modular, allowing for easy adjustment of the number of stages and depth. It is effective in preserving HR components and generating high-quality images, even with shallow networks. The results show that DBPN is capable of reconstructing HR images with high fidelity, especially for large scaling factors. The network's ability to incorporate feedback and dense connections makes it a powerful approach for SR tasks.