Edge Intelligence for Real-Time Image Recognition: A Lightweight Neural Scheduler Via Using Execution-Time Signatures on Heterogeneous Edge Devices

Abstract

Mobile Edge Computing (MEC) has emerged as a pivotal paradigm for enabling low-latency, high-efficiency AI applications at the network edge. However, heterogeneous hardware constraints across edge devices pose significant challenges in dynamic task scheduling. This paper introduces a novel neural execution-time signature model that learns hardware- and load-aware patterns to predict optimal device assignment for image recognition tasks. Leveraging a rich dataset collected from diverse platforms, including MacBook Pro, Raspberry Pi, and virtual machines, this research trains lightweight Artificial Neural Networks (ANNs) and Multilayer Perceptron (MLPs) to classify execution efficiency with over 99% accuracy. This research model captures subtle temporal and computational load features, enabling a smart scheduler that minimizes latency while maximizing resource utilization. Experimental results demonstrate superior performance in both balanced and imbalanced label scenarios, highlighting the model’s robustness and scalability. This work bridges predictive analytics and edge orchestration, offering a practical blueprint for next-generation edge intelligence systems. This research introduces a novel, lightweight neural scheduler that uses execution-time signatures and engineered temporal and performance features to predict task efficiency with over 99% accuracy across diverse edge devices like Raspberry Pi and MacBook Pro. In addition, it enables real-time, hardware-agnostic scheduling without per-device calibration, offering a scalable, cloud-independent solution for next-generation edge intelligence systems.