LarvaNet: Hierarchical Super-Resolution via Multi-Exit Architecture

Abstract

In recent years, image super-resolution (SR) methods using convolutional neural networks (CNNs) have achieved successful results. Nevertheless, it is often difficult to apply them in resource-constrained environments due to the requirement of heavy computation and huge storage capacity. To address this issue, we propose an efficient network model for SR, called LarvaNet. First, we investigate a number of architectural factors for a baseline model and find optimal settings in terms of performance, number of parameters, and running time. Based on that, we design our model using a multi-exit architecture. Our experiments show that the proposed method achieves state-of-the-art SR performance with a reasonable number of parameters and running time. We also show that the multi-exit architecture of the proposed model allows us to control the trade-off between resource consumption and SR performance by selecting which exit point to be used.