This paper provides a comprehensive survey of recent advances in image super-resolution (SR) using deep learning. Image SR aims to recover high-resolution (HR) images from low-resolution (LR) images, which is a challenging and ill-posed problem due to the multiple possible HR images corresponding to a single LR image. Traditional SR methods include prediction-based, edge-based, statistical, patch-based, and sparse representation approaches. However, deep learning has significantly improved SR performance, with methods ranging from early CNN-based approaches like SRCNN to recent GAN-based methods like SRGAN. The paper categorizes SR techniques into supervised, unsupervised, and domain-specific approaches. It covers benchmark datasets, performance metrics, and challenges in SR. Key metrics include PSNR, SSIM, and MOS, which evaluate reconstruction quality from different perspectives. Learning-based perceptual quality assessment methods, such as Ma and NIMA, aim to better capture human visual perception. The paper discusses various SR frameworks, including pre-upsampling, post-upsampling, progressive upsampling, and iterative up-and-down sampling. These frameworks differ in how they handle upsampling and the complexity of the models. Upsampling methods include traditional interpolation techniques and learnable layers like transposed convolution and sub-pixel layers. Network design strategies such as residual learning, recursive learning, multi-path learning, dense connections, attention mechanisms, and advanced convolutions are explored. These strategies enhance model performance by improving feature extraction, learning ability, and representation. The paper also highlights challenges and future directions in SR, including the trade-off between perceptual quality and distortion, and the need for efficient models for real-world applications.This paper provides a comprehensive survey of recent advances in image super-resolution (SR) using deep learning. Image SR aims to recover high-resolution (HR) images from low-resolution (LR) images, which is a challenging and ill-posed problem due to the multiple possible HR images corresponding to a single LR image. Traditional SR methods include prediction-based, edge-based, statistical, patch-based, and sparse representation approaches. However, deep learning has significantly improved SR performance, with methods ranging from early CNN-based approaches like SRCNN to recent GAN-based methods like SRGAN. The paper categorizes SR techniques into supervised, unsupervised, and domain-specific approaches. It covers benchmark datasets, performance metrics, and challenges in SR. Key metrics include PSNR, SSIM, and MOS, which evaluate reconstruction quality from different perspectives. Learning-based perceptual quality assessment methods, such as Ma and NIMA, aim to better capture human visual perception. The paper discusses various SR frameworks, including pre-upsampling, post-upsampling, progressive upsampling, and iterative up-and-down sampling. These frameworks differ in how they handle upsampling and the complexity of the models. Upsampling methods include traditional interpolation techniques and learnable layers like transposed convolution and sub-pixel layers. Network design strategies such as residual learning, recursive learning, multi-path learning, dense connections, attention mechanisms, and advanced convolutions are explored. These strategies enhance model performance by improving feature extraction, learning ability, and representation. The paper also highlights challenges and future directions in SR, including the trade-off between perceptual quality and distortion, and the need for efficient models for real-world applications.