SRCNN-based image transmission for autonomous vehicles in limited network areas

Abstract

High-quality images are crucial for navigation, obstacle detection, and environmental understanding, but transmitting high-resolution images over constrained networks presents significant challenges. This study introduces an image transmission system using super-resolution convolutional neural networks (SRCNN) to enhance image quality without increasing bandwidth requirements by transmitting low-resolution images and upscaling them with SRCNN. The first phase of the research involved data collection, in which information was acquired directly from an appropriate locus to produce training, validation, and testing datasets. The second, three SRCNN models (915, 935, and 955) were trained using such a training dataset. The last was an evaluation, in which model 915 showed quick learning and stable performance with initial high loss, while model 935 had rapid convergence but potential overfitting. Model 955 achieved high initial performance. Three SRCNN model configurations were tailored to the specific needs of autonomous electric vehicles operating in limited areas, such as the locus. Input image resolution ranged from 128×128 pixels to 256×256 pixels, while output resolution varied from 256×256 pixels to 512×512 pixels. These resolutions can be acceptable for efficient image transmission over IEEE 802.11ac, but on the long range (LoRa) network, it still produces some delay.