This paper introduces FSRCNN, a fast super-resolution convolutional neural network (SRCNN) that significantly improves the speed of image super-resolution while maintaining or enhancing the restoration quality. The original SRCNN, although effective, is computationally expensive, making it unsuitable for real-time applications. The authors identify two main limitations: the need for bicubic interpolation of the low-resolution (LR) image before processing and the high computational cost of the non-linear mapping step. To address these, they propose a compact hourglass-shaped CNN structure. The key innovations include replacing bicubic interpolation with a deconvolution layer, which allows direct learning from the original LR image to the high-resolution (HR) image, and restructuring the mapping layer to reduce computational complexity by shrinking and expanding feature dimensions. Additionally, they use smaller filter sizes and more mapping layers to enhance performance. The proposed model achieves a speed-up of over 40 times compared to the original SRCNN while maintaining superior restoration quality. They also present a smaller version, FSRCNN-s, which runs in real-time on a generic CPU. The model's architecture allows for efficient training and testing across different upscaling factors by sharing convolution layers, with only the deconvolution layer needing adjustment. Experiments show that FSRCNN outperforms existing methods in both speed and quality, making it suitable for real-time applications.This paper introduces FSRCNN, a fast super-resolution convolutional neural network (SRCNN) that significantly improves the speed of image super-resolution while maintaining or enhancing the restoration quality. The original SRCNN, although effective, is computationally expensive, making it unsuitable for real-time applications. The authors identify two main limitations: the need for bicubic interpolation of the low-resolution (LR) image before processing and the high computational cost of the non-linear mapping step. To address these, they propose a compact hourglass-shaped CNN structure. The key innovations include replacing bicubic interpolation with a deconvolution layer, which allows direct learning from the original LR image to the high-resolution (HR) image, and restructuring the mapping layer to reduce computational complexity by shrinking and expanding feature dimensions. Additionally, they use smaller filter sizes and more mapping layers to enhance performance. The proposed model achieves a speed-up of over 40 times compared to the original SRCNN while maintaining superior restoration quality. They also present a smaller version, FSRCNN-s, which runs in real-time on a generic CPU. The model's architecture allows for efficient training and testing across different upscaling factors by sharing convolution layers, with only the deconvolution layer needing adjustment. Experiments show that FSRCNN outperforms existing methods in both speed and quality, making it suitable for real-time applications.