DNN Model Partition and Resolution Control for Edge-Assisted Mobile Vision AI

초록

With the advancement of mobile device and network performance, the industry for mobile vision applications, such as image classification and object detection, is rapidly growing. When designing mobile vision applications, employing mobile edge computing (MEC) based on deep learning model partitioning can significantly improve the quality of user experience (QoE). Research on deep learning model partitioning has advanced in various directions over time, but until now, there has been no study that simultaneously partitionizes deep learning models by layers and adjusts the input frame size. Our proposed algorithm, Parecon, is based on Lyapunov optimization and dynamically adjusts 1) the number of frames to be processed, 2) model partition points, and 3) input frame size at each time slot. The proposed algorithm simultaneously optimizes processed fps, E2E latency, and top-1 accuracy. Through simulations, we confirmed that Parecon achieves a significant improvement in fps compared to existing algorithms while maintaining simular E2E latency and top-1 accuracy. © 2025, Korean Institute of Communications and Information Sciences. All rights reserved.

키워드