计算机与现代化 ›› 2024, Vol. 0 ›› Issue (11): 91-98.doi: 10.3969/j.issn.1006-2475.2024.11.014

• 图像处理 • 上一篇    下一篇

YOLOLW:一个新的轻量级目标检测模型




 
  

  1. (1.沈阳化工大学计算机科学与技术学院,辽宁 沈阳 110142; 2.辽宁省化工过程工业智能化技术重点实验室,辽宁 沈阳 110142)
  • 出版日期:2024-11-29 发布日期:2024-12-10
  • 基金资助:
    辽宁省自然科学基金资助项目(2022-MS-291); 辽宁省教育厅基本科研项目(LJKMZ20220781, LJKMZ20220783,
    LJKQZ20222457)

YOLOLW: A Novel Lightweight Object Detection Model

  1. (1. School of Computer Science and Technology, Shenyang University of Chemical Technology, Shenyang 110142, China; 
    2. Liaoning Provincial Key Laboratory of Intelligent Technology for Chemical Process Industry, Shenyang 110142, China)
  • Online:2024-11-29 Published:2024-12-10

摘要: 要满足日益增长的实时移动目标检测部署需求,目前的YOLO骨干网络仍存在许多不足。为此,本文提出基于锚框的轻量级目标检测模型YOLOLW。首先,它包含一个新颖的轻量级解耦头,以增强对分类和回归任务的关注,提高模型的准确性;其次,它设计一个轻量化和重参数化的网络结构,在保持其轻量化特性的同时,实现优异的检测精度;再次,通过动态卷积和跨层次关联有效整合浅层特征,增强特征金字塔结构(FPN);最后,引入空间注意机制和通道注意机制,显著提高了模型的准确性。实验结果验证了该模型的有效性。

关键词: 目标检测, YOLOLW, 轻量化模型, 注意力机制, FPN

Abstract:  In response to the growing demand for real-time mobile object detection deployment, the current YOLO backbone network falls short. Hence, this paper proposes YOLOLW, a lightweight object detection model based on the anchor frame. Firstly, it incorporates a novel lightweight decoupling header to enhance focus on classification and regression tasks and improve model accuracy. Secondly, it designs a lightweight and reparameterized network structure that achieves superior detection accuracy while maintaining its lightweight nature. Thirdly, it enhances the feature pyramid structure (FPN) by effectively integrating shallow features through dynamic convolution and cross-hierarchy association. Lastly, spatial and channel attention mechanisms are introduced to significantly boost the model’s accuracy. Experimental results validate the effectiveness of the YOLOLW model.

Key words: target detection, YOLOLW, lightweight model, attention mechanism, FPN

中图分类号: