Scholarlink

<p>Image dehazing is a representative research topic in the field of computer vision. In recent years, With the burgeoning advancement of deep learning, dehazing networks based on convolutional neural network (CNN) and Attention mechanisms are gaining widespread use. To address the limitations of existing dehazing networks in integrating CNNs and attention mechanisms, this paper proposes a dehazing network based on CNN and multiple Attention fusion mechanism. The proposed network first extracts and fuses multi-scale image features through a feature extraction network. Subsequently, it performs comprehensive dehazing at the channel, pixel, and spatial levels using the multiple attention fusion module. Meanwhile, the module proposed in this paper combines local residual and global residual structures, thereby notably enhancing the training speed and generalization performance of the model. Experiments demonstrate that the model proposed in this paper outperforms most current advanced dehazing algorithms and outperforms FFA-Net and AECR-Net on RESIDE-IN dataset with a PSNR of 38.67 dB and a SSIM of 0.990. Finally, the rationality of the module design is verified in the ablation study.</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

Single image dehazing based on convolutional neural network and multiple attention fusion

  • 期刊论文

  • 发表日期：

    2025年2月15日

查看全文
  • Jinyi Ming,
  • Zhidan Cai,
  • Shirong Li,
  • Sikai Bi,
  • Yongxin Ning

摘要

Image dehazing is a representative research topic in the field of computer vision. In recent years, With the burgeoning advancement of deep learning, dehazing networks based on convolutional neural network (CNN) and Attention mechanisms are gaining widespread use. To address the limitations of existing dehazing networks in integrating CNNs and attention mechanisms, this paper proposes a dehazing network based on CNN and multiple Attention fusion mechanism. The proposed network first extracts and fuses multi-scale image features through a feature extraction network. Subsequently, it performs comprehensive dehazing at the channel, pixel, and spatial levels using the multiple attention fusion module. Meanwhile, the module proposed in this paper combines local residual and global residual structures, thereby notably enhancing the training speed and generalization performance of the model. Experiments demonstrate that the model proposed in this paper outperforms most current advanced dehazing algorithms and outperforms FFA-Net and AECR-Net on RESIDE-IN dataset with a PSNR of 38.67 dB and a SSIM of 0.990. Finally, the rationality of the module design is verified in the ablation study.