Scholarlink

<p>In adverse weather conditions such as haze, a large number of suspended particles are present in the atmosphere, which may interfere with the real-time and high-quality acquisition of visual images in applications such as remote sensing, monitoring, and automotive systems. Existing neural network-based algorithms require high computational power and storage, resulting in poor real-time performance. To address these issues, this paper proposes a lightweight spanning fusion attention convolutional neural network (SFA-Net), which adopts an autoencoder architecture and utilizes feature attention fusion to manipulate the interaction between Channel Attention (CA) and Pixel Attention (PA) information, and connects the feature information of the front and back convolutional layers through an adaptive mixup fusion module to train the defogging model and reconstruct clear images. In addition, a configurable and efficient hardware acceleration circuit based on the network is designed, which can be applied to different indoor and outdoor scenarios. Experimental results show that the proposed method has a good dehazing ability with only 720 parameters, and is better than the indicators of other hardware implementation algorithms in the comparison of objective data sets. In terms of hardware, compared with the brightness improved by light-dehazeNet (BILD-Net), the resource is reduced by 48%, and the processing speed can reach a maximum frame rate of 120 frames per second, which meets the requirements of real-time processing.</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 high-speed hardware accelerator for lightweight dehazing neural network based on SFA-Net

  • 期刊论文

  • 发表日期：

    2025年4月15日

查看全文
  • Zhe Chen,
  • Gaoming Du,
  • Zhenmin Li,
  • Xiaolei Wang,
  • Yongsheng Yin

摘要

In adverse weather conditions such as haze, a large number of suspended particles are present in the atmosphere, which may interfere with the real-time and high-quality acquisition of visual images in applications such as remote sensing, monitoring, and automotive systems. Existing neural network-based algorithms require high computational power and storage, resulting in poor real-time performance. To address these issues, this paper proposes a lightweight spanning fusion attention convolutional neural network (SFA-Net), which adopts an autoencoder architecture and utilizes feature attention fusion to manipulate the interaction between Channel Attention (CA) and Pixel Attention (PA) information, and connects the feature information of the front and back convolutional layers through an adaptive mixup fusion module to train the defogging model and reconstruct clear images. In addition, a configurable and efficient hardware acceleration circuit based on the network is designed, which can be applied to different indoor and outdoor scenarios. Experimental results show that the proposed method has a good dehazing ability with only 720 parameters, and is better than the indicators of other hardware implementation algorithms in the comparison of objective data sets. In terms of hardware, compared with the brightness improved by light-dehazeNet (BILD-Net), the resource is reduced by 48%, and the processing speed can reach a maximum frame rate of 120 frames per second, which meets the requirements of real-time processing.