LightSRNet A Lightweight, Quantized Super-Resolution Network Optimized for Embedded IoT Applications

Abstract High-resolution imaging would still play an important role in numerous practical problems like smart surveillance, remote sensing as well as autonomous navigation. Deep super-resolution (SR) models are difficult to deploy to such resource-constrained edge devices because of power constraints, memory capacities, and limited computational capacity. This paper introduces LightSRNet a new light-weight network on the topic of super-resolution networks that optimize efficiency in image reconstruction on embedded processing and IoT systems. The network combines effective building blocks like depthwise separable convolutions, residual feature distillation, and attention mechanisms to make it simpler yet retain the quality of the images. In order to further improve deployment efficiency, the model uses both post-training quantization (PTQ) and quantization-aware training (QAT), including 8-bit and 4-bit precision types. These techniques reduce model size and latency of inference significantly with minor changes in accuracy. It uses a hybrid loss term containing both pixel-based and perceptual ones, the former ensuring structural faithfulness, while the latter structural recovery and texture. The model shows a lot of efficiency in terms of trading performance and computational elegance through its extensive experiments performed on some standard benchmarking like DIV2K and Urban100. Also, the quantized variants of LightSRNet demonstrate competitive values of PSNR and SSIM scales and provide a much lower cost of power and latency using the hardware like the NVIDIA Jetson Nano. The effectiveness of individual elements of the architectures was also confirmed by an ablation. An end-to-end pipeline with ONNX Runtime shows how to work in real-time in a real-world setting. This study proposes a workable and scalable solution to high-quality image enhancement when dealing with edge AI.