Quantized Neural Network Architecture for Hardware Efficient Real-Time 4K Image Super-Resolution
George Joseph, E. P. Jayakumar · 2024
Super-resolution refers to a class of techniques for producing high-resolution images from their low-resolution counterparts, using computational algorithms and methods. The higher the resolution, the greater the detail in the image, which is critical for current digital imaging applications. The techniques can either be classical or deep learning (DL) based methods like convolutional neural network (CNN). Although the CNN-based methods provide better results in terms of image reconstruction quality, the computational complexity due to the increasing number of parameters makes it difficult to provide real-time performance without the additional cost of resource utilization. In this paper, we propose optimizations to the fast super-resolution convolutional neural network (FSRCNN), a CNN-based super-resolution method by applying suitable quantization and modifications, for reducing the computational cost as well as memory footprint and implementing a hardware-efficient design on FPGA through FINN. In addition to achieving nearly a 50% reduction in resource utilization, our design enables real-time ×2 image super-resolution from 1080p HD to 4K UHD with no significant loss in image quality compared to the original FSRCNN network.