Optimizing Deep Image Super-Resolution Techniques for Low-Power Devices

Abstract

Over the past couple years, deep learning Image Super-Resolution (SR) has taken large steps that arrive at artifacts-free image reconstructions provided that the input is low-resolution. However, implementations of the models to low-power devices, viz. smart phones, drones, and embedded systems are problematic due to the limited available resources that pose daunting challenges of varying levels of complexity. The present research will contain the development of deep image super-resolution method that can be applied most exquisitely in the context of low power. The paper evaluates and compares the efficiencies in the terms of lightweight of convolutional neural networks (CNNs), quantization, model pruning and knowledge distillation strategies as the method of simplifying a model and, at the same time, maintain the quality of the images. The experimental results depict that well-generated lightweight SR model could be as competitive as a full-scale in terms of competing with a large reduction of resources. The paper proposes a realistic model of executing SR models on edge apparatus, which promotes cost saving and convenient systems to enhance images.