Scholarlink

<p>Infrared small target detection plays an important role in infrared search and tracking applications. In recent years, deep learning techniques have been introduced to this task with remarkable results. However, the current method is limited by reasons such as network structure and feature extraction capability, which makes it difficult to complete the detection of deep infrared small targets and infrared small targets in complex backgrounds. To address the above problems, this paper proposes attention-guided dual feature extraction infrared small target detection network, which consists of two levels of feature extraction, a multi-scale feature extraction mechanism and a contextual feature extraction mechanism, respectively. The multi-scale feature extraction mechanism implements the feature extraction operation of the first layer, and by fusing and splicing the features of different layers, the information interaction between neighboring layers is achieved, which can effectively prevent the loss of small targets in the deep layer. Among them, the target attention mechanism enhances the model’s attention to small targets and provides effective attention allocation for the detection of small infrared targets. And the multi-resolution feature interaction mechanism can better preserve the detail information of the image and enhance the model’s ability to perceive small targets. The contextual feature extraction mechanism implements the second-level feature extraction operation, which is able to accurately locate the position of small targets in complex background environments by effectively integrating global semantics and local semantics. The experimental results show that the method in this paper achieves 99.81% vs. 87.58% Precision, 98.96% vs. 88.12% Recall,99.38% vs. 87.85% F-measure and 98.77% vs. 78.33% mIoU on NUDT-SIRST and SIRST Aug datasets, respectively, which is better than the other current state-of-the-art methods. Training code and pretrained models are available at <a href="https://github.com/wwwyangxiaoyucom/new_network">https://github.com/wwwyangxiaoyucom/new_network</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

Attention-guided dual feature extraction approach for small target detection in infrared images

  • 期刊论文

  • 发表日期：

    2025年1月8日

查看全文
  • Yufang Yang,
  • Yining Xie,
  • Jun Cao,
  • Kaihua Yang

摘要

Infrared small target detection plays an important role in infrared search and tracking applications. In recent years, deep learning techniques have been introduced to this task with remarkable results. However, the current method is limited by reasons such as network structure and feature extraction capability, which makes it difficult to complete the detection of deep infrared small targets and infrared small targets in complex backgrounds. To address the above problems, this paper proposes attention-guided dual feature extraction infrared small target detection network, which consists of two levels of feature extraction, a multi-scale feature extraction mechanism and a contextual feature extraction mechanism, respectively. The multi-scale feature extraction mechanism implements the feature extraction operation of the first layer, and by fusing and splicing the features of different layers, the information interaction between neighboring layers is achieved, which can effectively prevent the loss of small targets in the deep layer. Among them, the target attention mechanism enhances the model’s attention to small targets and provides effective attention allocation for the detection of small infrared targets. And the multi-resolution feature interaction mechanism can better preserve the detail information of the image and enhance the model’s ability to perceive small targets. The contextual feature extraction mechanism implements the second-level feature extraction operation, which is able to accurately locate the position of small targets in complex background environments by effectively integrating global semantics and local semantics. The experimental results show that the method in this paper achieves 99.81% vs. 87.58% Precision, 98.96% vs. 88.12% Recall,99.38% vs. 87.85% F-measure and 98.77% vs. 78.33% mIoU on NUDT-SIRST and SIRST Aug datasets, respectively, which is better than the other current state-of-the-art methods. Training code and pretrained models are available at https://github.com/wwwyangxiaoyucom/new_network.