Lightweight CNN Architecture for Real-Time Image Super-Resolution in Edge Devices

Abstract

Image super resolution (SR) is a fundamental problem in computer vision that has a wide range of applications in both medical imaging, video surveillance and mobile photography. While high-accuracy SR models can be deployed to end devices, the lack of computational power, memory and energy consumption constraints prevents them from being deployed in edge devices. This work proposes a novel lightweight convolutional neural network (CNN) architecture, EdgeSRNet, for real time image super resolution on low power embedded platforms. To achieve this, depthwise separable convolutions, residual efficient building blocks, along with sub pixel convolutional layers are incorporated to significantly reduce model complexity while maintaining high reconstruction fidelity. We optimize the architecture for low latency inference, with real time inference achievable without GPU acceleration. Extensive experiments have been conducted to evaluate its effectiveness on publicly available benchmark of Set5, Set14, and BSD100. Results show that EdgeSRNet provides competitive PSNR/SSIM performance with fewer than 500K parameters and under 1.5 GFLOPs per forward pass. Further, we compare our EdgeSRNet with several existing lightweight SR models on edge devices, such as Raspberry Pi 4 and NVIDIA Jetson Nano, and show that the EdgeSRNet achieves better visual quality and high computational efficiency than existing lightweight models on edge devices. With these attributes, EdgeSRNet shows great potential for edge real time image enhancement in resource constrained scenarios, for instance IoT devices, smart cameras, autonomous systems, and mobile platforms.