计算机工程与应用 ›› 2025, Vol. 61 ›› Issue (8): 226-238.DOI: 10.3778/j.issn.1002-8331.2312-0134

• 图形图像处理 • 上一篇    下一篇

结合特征融合注意力的规范化卷积图像去雾网络

王燕,卢鹏屹,他雪   

  1. 兰州理工大学 计算机与通信学院,兰州 730050
  • 出版日期:2025-04-15 发布日期:2025-04-15

Normalized Convolutional Image Dehazing Network Combined with Feature Fusion Attention

WANG Yan, LU Pengyi, TA Xue   

  1. School of Computer and Communication, Lanzhou University of Technology, Lanzhou 730050, China
  • Online:2025-04-15 Published:2025-04-15

摘要: 图像去雾方法可以从受雾、霾或大气干扰影响的模糊图像中恢复清晰图像,在自动驾驶、监控系统等领域有重要意义。近年来,基于深度学习方法在图像去雾领域取得了显著进展,但在去雾过程中会损失一些细节信息,导致纹理信息恢复不足、去雾不均匀现象。为了解决这个问题,提出了一种端到端的单幅图像去雾方法,称为EFAN-Net。该方法由编码器模块、去雾模块、解码器模块组成。编码器模块用于提取图像特征信息传递给去雾模块;去雾模块通过特征融合组(FFG)获得更多图像信息使去雾图像颜色失真更小、伪影更少,将获得的图像信息传递给深度规范化修正卷积块(DNCC)减少协变量偏移,使模型更容易训练。多路径特征卷积块(MFCB)获得纹理细节更丰富的图像信息,最后经过解码器模块通过反卷积和上采样操作将低维的特征映射转换回高维的原始输入空间,得到去雾图像。大量实验结果表明,所提方法在定量和定性上都取得了较好的结果并优于相关的最新方法。

关键词: 图像去雾, 深度学习, 编码器-解码器, 深度规范化修正卷积块

Abstract: Image dehazing methods can recover clear images from blurred images affected by fog, haze or atmospheric disturbances, which is of great significance in the fields of automatic driving, surveillance systems, and so on. In recent years, significant progress has been made in the field of image dehazing based on deep learning methods, but some detail information is lost during the dehazing process, resulting in insufficient recovery of texture information and uneven dehazing phenomenon. In order to solve this problem, an end-to-end single image dehazing method, called EFAN-Net, is proposed, which consists of an encoder module, a dehazing module, and a decoder module. The encoder module is used to extract the image feature information to be passed to the dehazing module; the dehazing module obtains more image information through the feature fusion group (FFG) to make the dehazed image with less color distortion and artifacts, and passes the obtained image information to the depth-normalized corrected convolution block (DNCC) to reduce covariate offsets so that the model can be trained more easily. The multipath feature convolution block (MFCB) obtains image information with richer texture details, and finally the decoder module converts the low-dimensional feature mapping back to the high-dimensional original input space through the inverse convolution and up-sampling operations to obtain the dehazed image. A large number of experimental results show that the proposed method achieves better results both quantitatively and qualitatively and outperforms the related state-of-the-art methods.

Key words: image dehazing, deep learning, encoder-decoder, deep normalization correction convolutional block