This paper proposes a fast and efficient Super-Resolution Convolutional Neural Network (FSRCNN) to accelerate the existing Super-Resolution Convolutional Neural Network (SRCNN) for image super-resolution. The main contributions include: 1) A compact hourglass-shaped CNN structure that enables end-to-end mapping between low-resolution (LR) and high-resolution (HR) images without preprocessing. 2) A significant speedup of more than 40 times compared to SRCNN-Ex while maintaining excellent performance. 3) A small version of FSRCNN (FSRCNN-s) that achieves real-time performance (24 fps) on a generic CPU with better restoration quality than SRCNN. 4) A transfer strategy that allows fast training and testing across different upscaling factors without loss of restoration quality. The FSRCNN structure is designed to address two key limitations of SRCNN: the need for bicubic interpolation and the costly non-linear mapping step. The first limitation is addressed by replacing bicubic interpolation with a deconvolution layer that directly maps the original LR image to the HR image. The second limitation is addressed by introducing shrinking and expanding layers to reduce the dimensionality of the feature space and decomposing the mapping layer into multiple smaller layers. These changes significantly reduce the computational complexity and improve the speed of the network. Experiments show that FSRCNN achieves a speedup of more than 40 times compared to SRCNN-Ex while maintaining superior performance. The FSRCNN-s version achieves real-time performance (24 fps) on a generic CPU with better restoration quality than SRCNN. Additionally, the FSRCNN structure allows for fast training and testing across different upscaling factors by sharing convolution layers among different upscaling factors. This makes FSRCNN a highly efficient and effective solution for image super-resolution.This paper proposes a fast and efficient Super-Resolution Convolutional Neural Network (FSRCNN) to accelerate the existing Super-Resolution Convolutional Neural Network (SRCNN) for image super-resolution. The main contributions include: 1) A compact hourglass-shaped CNN structure that enables end-to-end mapping between low-resolution (LR) and high-resolution (HR) images without preprocessing. 2) A significant speedup of more than 40 times compared to SRCNN-Ex while maintaining excellent performance. 3) A small version of FSRCNN (FSRCNN-s) that achieves real-time performance (24 fps) on a generic CPU with better restoration quality than SRCNN. 4) A transfer strategy that allows fast training and testing across different upscaling factors without loss of restoration quality. The FSRCNN structure is designed to address two key limitations of SRCNN: the need for bicubic interpolation and the costly non-linear mapping step. The first limitation is addressed by replacing bicubic interpolation with a deconvolution layer that directly maps the original LR image to the HR image. The second limitation is addressed by introducing shrinking and expanding layers to reduce the dimensionality of the feature space and decomposing the mapping layer into multiple smaller layers. These changes significantly reduce the computational complexity and improve the speed of the network. Experiments show that FSRCNN achieves a speedup of more than 40 times compared to SRCNN-Ex while maintaining superior performance. The FSRCNN-s version achieves real-time performance (24 fps) on a generic CPU with better restoration quality than SRCNN. Additionally, the FSRCNN structure allows for fast training and testing across different upscaling factors by sharing convolution layers among different upscaling factors. This makes FSRCNN a highly efficient and effective solution for image super-resolution.