A Study on Image Resolution and Object Scale Adjustment for Efficient Object Detection in Mobile Network Environments

Abstract

Object detection is one of the most fundamental and core research areas in the field of computer vision, and the YOLO Series, a representative model series, is widely utilized across various artificial intelligence systems. Mobile networks serve as a crucial connectivity element that links nearly all industrial sectors, connecting various IoT devices through these networks. A typical example is network cameras (CCTV). Some deep learning AI models often exhibit degraded object detection performance compared to their reported benchmark results. While multiple factors may contribute to this, one well-known reason is the difference in characteristics between publicly available training datasets and images collected in CCTV environments. Due to the inherent bandwidth limitations of wireless networks, data transmission is often constrained. Particularly in mobile network environments, various approaches such as applying edge computing have been researched to reduce network load for object detection models deployed on CCTV systems. In this study, we systematically and linearly adjusted image resolution and object scale in video/image data transmitted over the network to analyze their impact on detection performance. Through this, the goal is to explore practical methods for achieving efficient real-time object detection that consider the constraints of network environments.