SlimDL: Deploying ultra-light deep learning model on sweeping robots

Advanced object detection methods have yielded impressive progress in recent years. However, the computational constraints of edge mobile devices present significant deployment challenges for state-of-the-art algorithms. We propose a deep learning deployment framework with two stages: model adaptati...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Engineering applications of artificial intelligence Jg. 149; S. 110415
Hauptverfasser: Sun, Xudong, Wang, Yu, Liu, Zhanglin, Gao, Shaoxuan, He, Wenbo, Tong, Chao
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier Ltd 01.06.2025
Schlagworte:
Model adaptation
Model compression
Object detection
Model compression
Model adaptation
Object detection
ISSN:0952-1976
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Advanced object detection methods have yielded impressive progress in recent years. However, the computational constraints of edge mobile devices present significant deployment challenges for state-of-the-art algorithms. We propose a deep learning deployment framework with two stages: model adaptation and compression. Our method enhance “You Only Look Once version 5” (YOLOv5) with lightweight modules, which improves detection performance while reducing computational load. Additionally, we present a pruning algorithm, employing adaptive batch normalization and iterative pruning. Our evaluation on “Microsoft Common Objects in Context” (MSCOCO) dataset and custom SweepRobot datasets demonstrates that our method consistently outperforms state-of-the-art approaches. On the SweepRobot dataset, our method doubled YOLOv5’s detection speed on the sweeping robot from 15.69 frames per second (FPS) to 30.77 FPS, maintaining 97.3% performance at 20% of the computational cost. Even on Graphics Processing Unit equipped devices, our method achieved 1.8% and 2.8% higher Average Precision compared to direct scaling and pruning with the original pruning algorithm.
ISSN:0952-1976
DOI:10.1016/j.engappai.2025.110415