Scholarlink

<p>In this paper, we propose a novel model compression approach for YOLOv8n - the lightweight variant of YOLOv8 - by applying Higher-Order Singular Value Decomposition (HOSVD) to its 2D convolutional layers. We introduce an efficient strategy to decompose and fine-tune the network, achieving significant parameter reduction with minimal performance loss. Experimental results on the COCO dataset demonstrate that our compressed model reduces the number of parameters by up to 50% , while maintaining high detection accuracy with only a 7.77% drop in mAP@0.5 . Furthermore, the proposed architecture improves inference speed by up to 17.86% , as measured by frames per second (FPS), showing its potential for real-time applications and future deployment on resource-constrained devices.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing YOLOv8n using hOSVD for efficient object detection: towards real-time video surveillance applications

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Alaa EL ICHI,
  • Wissam Kaddah,
  • Marwa EL Bouz ,
  • Isabelle Badoc

摘要

In this paper, we propose a novel model compression approach for YOLOv8n - the lightweight variant of YOLOv8 - by applying Higher-Order Singular Value Decomposition (HOSVD) to its 2D convolutional layers. We introduce an efficient strategy to decompose and fine-tune the network, achieving significant parameter reduction with minimal performance loss. Experimental results on the COCO dataset demonstrate that our compressed model reduces the number of parameters by up to 50% , while maintaining high detection accuracy with only a 7.77% drop in mAP@0.5 . Furthermore, the proposed architecture improves inference speed by up to 17.86% , as measured by frames per second (FPS), showing its potential for real-time applications and future deployment on resource-constrained devices.