Scholarlink

<p>Object detection based on images captured by UAVs has become a hot research topic. However, images have a rich variety of small targets and complex backgrounds. This poses significant challenges for existing object detectors. Furthermore, due to the limitations of UAV platforms, it is difficult to deploy complicated models. Therefore, a novel lightweight detection network LDN-UAV is introduced in this work. Firstly, the YOLOv5s network is redesigned to obtain the YOLOv5ss basic network, which reduces the complexity of the model while improving the detection performance. Moreover, a lightweight feature enhancement module is devised, which enhances the spatial utilization of features. Next, we propose a novel lightweight convolutional operation that maps receptive-field features to a specific size via a shared MLP, which simplifies the model and improves the performance. Moreover, a context augmentation module is created, which aggregates contextual information to increase the benefits of features through MLP branching and Softmax. Finally, a lightweight Decoupled-Head is designed to ensure efficient performance of the detection. To validate the advantages of proposed lightweight convolutions, we conduct extensive experiments on COCO2017 and VOC 7+12. Additionally, for images captured by UAVs, relevant experiments are performed based on VisDrone-DET2021. The results of all experiments demonstrate that the proposed method achieves better detection performance compared to state-of-the-art lightweight detectors. Compared to the original baseline model, LDN-UAV uses only 2.46M parameters and increased the metrics mAP50 and mAP by 8.6 % and 5.9 %, respectively. Code is available at <a href="https://github.com/CV-ZhangXin/LDN-UAV">https://github.com/CV-ZhangXin/LDN-UAV</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 object detector based on changeable-size lightweight convolution and context augmentation module for images captured by UAVs

  • 期刊论文

  • 发表日期：

    2024年12月31日

查看全文
  • Xin Zhang,
  • Degang Yang,
  • Tingting Song,
  • Yichen Ye,
  • Yingze Song,
  • Jie Zhou,
  • Jie Chen

摘要

Object detection based on images captured by UAVs has become a hot research topic. However, images have a rich variety of small targets and complex backgrounds. This poses significant challenges for existing object detectors. Furthermore, due to the limitations of UAV platforms, it is difficult to deploy complicated models. Therefore, a novel lightweight detection network LDN-UAV is introduced in this work. Firstly, the YOLOv5s network is redesigned to obtain the YOLOv5ss basic network, which reduces the complexity of the model while improving the detection performance. Moreover, a lightweight feature enhancement module is devised, which enhances the spatial utilization of features. Next, we propose a novel lightweight convolutional operation that maps receptive-field features to a specific size via a shared MLP, which simplifies the model and improves the performance. Moreover, a context augmentation module is created, which aggregates contextual information to increase the benefits of features through MLP branching and Softmax. Finally, a lightweight Decoupled-Head is designed to ensure efficient performance of the detection. To validate the advantages of proposed lightweight convolutions, we conduct extensive experiments on COCO2017 and VOC 7+12. Additionally, for images captured by UAVs, relevant experiments are performed based on VisDrone-DET2021. The results of all experiments demonstrate that the proposed method achieves better detection performance compared to state-of-the-art lightweight detectors. Compared to the original baseline model, LDN-UAV uses only 2.46M parameters and increased the metrics mAP50 and mAP by 8.6 % and 5.9 %, respectively. Code is available at https://github.com/CV-ZhangXin/LDN-UAV.