Robust Bare-Bone CNN Applying for Tactical Mobile Edge Devices

Abstract

Artificial intelligence (AI) technologies such as image recognition, classification, and generative AI are constantly evolving rapidly. Many of these techniques operate in high-performance computing environments because they use complex architectural models and millions of parameters to improve inference and prediction performance. In recent years, there has been a growing demand for AI applications in the defense domain, and considerable research has been conducted. In tactical environments, images are used for various functions, such as creating an operational overlay and analyzing information, and AI can be leveraged for these functions. However, tactical mobile edge devices have insufficient computing resources, which limits their ability to simultaneously perform various applications such as service-proven command and control, intelligence analysis, and fire operations, as well as applications using existing AI models. Therefore, a robust bare-bone convolutional neural network (CNN) model that can support reliable services on tactical mobile edge devices was proposed. The proposed model uses only four convolutional layers, has a performance equivalent to or better than that of existing CNN models, and has a stable and sufficient performance potential to perform multiple applications simultaneously. For experimental validation of the proposed model, a military symbol inferencer using self-collected handwritten military symbol images was implemented. This inferencer has an average accuracy of 95.42% when used alone, with CPU utilization reduced by up to 31.3% and inference time reduced by up to 47.2%. When running multiple applications in parallel, CPU utilization was reduced by up to 23.7% and inference time by up to 55.9%.