Scholarlink

<p>To address the challenges of excessive model parameters and limited detection accuracy in small object detection for UAV imagery, this paper proposed a Lightweight Small Object Detection (LWSDet) model. Firstly, an adaptive feature selection module is designed. By using a gating mechanism and depthwise separable convolution to dynamically weight multiscale features, this module can reduce redundant parameters while enhancing the representation ability of key features. Secondly, a multilayer feature fusion module is proposed. By fusing adjacent features, it can effectively integrate the detailed information in the shallow layers with the semantic context in the deep layers, thereby improving the detection robustness in complex scenarios. Finally, a partial ghost convolution module is introduced. By improving GhostConv, this module can retain more detailed features while reducing the number of parameters. Experimental results on three public datasets demonstrate that LWSDet achieves a 2.9% improvement in mean average precision (mAP50) and a 36.8% reduction in model parameters compared to the baseline. This code is available at <a href="https://github.com/wujie-max/LWSDet.git">https://github.com/wujie-max/LWSDet.git</a>.</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

A lightweight adaptive feature selection network for enhanced small object detection in UAV imagery

  • 期刊论文

  • 发表日期：

    2025年6月25日

查看全文
  • Jinxia Yu,
  • Jie Wu,
  • Yongli Tang

摘要

To address the challenges of excessive model parameters and limited detection accuracy in small object detection for UAV imagery, this paper proposed a Lightweight Small Object Detection (LWSDet) model. Firstly, an adaptive feature selection module is designed. By using a gating mechanism and depthwise separable convolution to dynamically weight multiscale features, this module can reduce redundant parameters while enhancing the representation ability of key features. Secondly, a multilayer feature fusion module is proposed. By fusing adjacent features, it can effectively integrate the detailed information in the shallow layers with the semantic context in the deep layers, thereby improving the detection robustness in complex scenarios. Finally, a partial ghost convolution module is introduced. By improving GhostConv, this module can retain more detailed features while reducing the number of parameters. Experimental results on three public datasets demonstrate that LWSDet achieves a 2.9% improvement in mean average precision (mAP50) and a 36.8% reduction in model parameters compared to the baseline. This code is available at https://github.com/wujie-max/LWSDet.git.