Edge AI-Enabled Backbone Optimization for Real-Time Object Detection in Computing Power Networks

Object detection is a fundamental task in computer vision with broad applications in autonomous transportation driving, smart surveillance, and traffic monitoring. In the context of Computing Power Networks (CPNs), which interconnect cloud, edge, and terminal nodes to support distributed AI services, deploying efficient object detection models under constrained computational resources is a critical challenge, particularly at the edge and terminal layers. This study investigates backbone optimization for the YOLOv11M object detection framework to improve computational efficiency while maintaining detection performance. We propose two lightweight variants, YOLOv11M-MN and YOLOv11M-Shuffle, by replacing the original backbone with MobileNetV3-Small and ShuffleNetV2, respectively. All Edge AI-enabled models share an identical detection head and training pipeline to ensure fair and controlled comparisons. To reflect resource-limited CPN environments, all experiments are conducted under CPU-only settings with staged training budgets. Performance is evaluated using the COCO128 dataset in terms of FLOPs, parameter count, inference latency, and detection accuracy. Experimental results demonstrate that the proposed lightweight backbones significantly reduce computational overhead and inference time, while exhibiting different accuracy–efficiency trade-offs, highlighting their suitability for selective deployment across heterogeneous CPN layers.