Efficient Image Transmission for Autonomous Systems Using Residual Dense Feature Networks Over LoRa Networks

Abstract

Autonomous systems face challenges in transmitting high-quality images over bandwidth-constrained networks like LoRa, which operates at data rates of 0.3–50 kbps. This study proposes the Residual Dense Feature Network (RDF Net), a super-resolution model designed to optimize image transmission within the constraints of LoRa networks. By leveraging Contrast-Aware Channel Attention (CCA), Enhanced Spatial Attention (ESA), Blueprint Separable Convolution (BSConv), and a progressive approach, RDF Net achieves 20x upscaling, enabling low-resolution images (40x40 pixels) to be reconstructed into high-resolution outputs (800x800 pixels) on a central server. Experimental evaluations demonstrate that Model-4, combining CCA and ESA, delivers state-of-the-art perceptual quality and structural fidelity, while Model-3, using ESA, offers a computationally efficient alternative for resource-constrained scenarios. Simulations of LoRa’s bandwidth limitations reveal that transmitting a single 40x40 image requires approximately 0.208–0.56 seconds at a data rate of 50 kbps. While this demonstrates the feasibility of near real-time communication, the trade-off between latency and visual fidelity remains a critical consideration, particularly for latency-sensitive applications. These findings underscore RDF Net’s potential to address the challenges of high-quality visual communication in bandwidth-constrained environments, paving the way for enhanced autonomous system applications. Further optimization, including adaptive compression strategies, and testing on actual LoRa hardware are recommended to validate its performance in real-world scenarios and explore its applicability to diverse autonomous systems.