Binary neural networks for image super-resolution

Abstract

Recently, deep convolutional neural networks (DNNs) have achieved state-of-the-art performance on the super-resolution task. However, bigger and deeper networks often lead to high computational cost and high memory usage, preventing massive applications on resource-limited devices. Therefore, reducing the model storage and computation costs will expand the application range of DNNs. Binary neural networks, which could reduce the model size and allow for efficient inference, are energy-efficient for embedded devices. It motivates us to apply the binary methods into super-resolution field, enabling us to meet the requirements of the hardware platforms in practical applications. Therefore, this paper focuses on the binary image super-resolution network. To this end, we first review the current binary neural networks and introduce the technical details and algorithm characteristics of the existing algorithms in detail. Then, we propose to apply the existing binary training methods to the super-resolution task. On this basis, we propose a new binary super-resolution method, which advances network performance by improving the representation ability of the quantized network in the forward propagation and reducing information loss in the backward propagation. Extensive experiments demonstrate that the proposed method not only outperforms existing binary methods but also outperforms the state-of-the-arts binary super-resolution methods.