A Hybrid Data Manipulation Approach for Energy and Latency-Efficient Vision-Aided UDNs

The combination of deep learning (DL) and computer vision (CV) is shaping the future of wireless communications by supporting the operations of ultra-dense networks (UDNs). However, vision-aided wireless communications (VAWC) are highly dependent on DL algorithms that rely on a wide range of multimodal data stored at a central location. Although the performance of the DL model is improved when the model becomes deeper, the need for a large number of datasets for model training incurs more computational complexity in terms of model training time and storage size. Hence, the energy efficiency of the network will become worse due to the higher energy costs associated with model training and transmitting a large amount of data over wireless links. Therefore, a crit-ical challenge is to reduce the computational complexity and bandwidth utilisation of DL-based vision-aided UDNs without compromising their performance. In this paper, we adopt single-channel (SICH) images, joint photographic expert group (JPEG) image compression (COMP), and object detection (ODET) to form a hybrid data manipulation technique. This technique can reduce the model computation cost and data storage volume, as well as alleviate the transmission burden on the wireless links to make future wireless networks more reliable and energy efficient. Specifically, this technique is used to manipulate datasets before using them in model training. Compared to reference datasets, simulation results show that our hybrid technique achieves the best performance in reducing the model computation by 34%, a significant reduction of 86% in memory size for data storage, reducing data transmission time by 83%, and 82.5% more energy efficient networks.