作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (8): 174-181. doi: 10.19678/j.issn.1000-3428.0064826

• 图形图像处理 • 上一篇    下一篇

基于高低频特征增强的图像去雾

卢昂1,2, 储珺1,3, 冷璐1,3   

  1. 1. 江西省图像处理与模式识别重点实验室(南昌航空大学), 南昌 330063
    2. 南昌航空大学 信息工程学院, 南昌 330063
    3. 南昌航空大学 软件学院, 南昌 330063
  • 收稿日期:2022-05-26 出版日期:2023-08-15 发布日期:2022-10-21
  • 作者简介:

    卢昂(1996—),男,硕士研究生,主研方向为计算机视觉、图像去雾

    储珺,教授、博士

    冷璐,教授、博士

  • 基金资助:
    国家自然科学基金(62162045); 国家自然科学基金(61866028); 江西省技术创新引导类计划项目(20212BDH81003); 南昌航空大学研究生创新基金(YC2020-033)

Image Dehazing Based on High-Frequency and Low-Frequency Feature Enhancement

Ang LU1,2, Jun CHU1,3, Lu LENG1,3   

  1. 1. Key Laboratory of Jiangxi Province for Image Processing and Pattern Recognition(Nanchang Hangkong University), Nanchang 330063, China
    2. School of Information Engineering, Nanchang Hangkong University, Nanchang 330063, China
    3. School of Software, Nanchang Hangkong University, Nanchang 330063, China
  • Received:2022-05-26 Online:2023-08-15 Published:2022-10-21

摘要:

图像去雾是典型的不适定问题,编解码网络是常用的去雾网络架构,编解码网络由编码器、解码器和连接两者的特征转换器构成。已有去雾算法生成的去雾图像通常质量较低,存在局部细节去雾不彻底、颜色失真或引入噪声等问题。针对基于编解码网络的去雾算法在设计特征转换器时没有充分利用小尺度特征、解码阶段仅利用对应层编码特征的问题,提出一种基于高低频特征增强的去雾算法。在特征转换阶段,设计扩张残差组件并组成上下文聚合网络,充分利用大感受野的低分辨率特征,提取特征图远距离相关性,增强低频小尺度特征。设计基于通道注意力的多级特征重用网络,实现浅层高频特征的重利用,同时深度融合解码重建特征,增强视觉感知特征的恢复。在编码阶段构建视觉特性感知模块,利用残差块在局部建模方面的优势增强浅层丰富的高频视觉特征。实验结果表明,与AOD-Net、PFF-Net等去雾算法相比,该算法的PSNR和SSIM指标均有明显优势,在室内合成数据集SOTS和室外真实数据集Dense-Haze上,所提算法的PSNR和SSIM分别高出性能次优算法0.77 dB、0.000 7和0.40 dB、0.037 1。

关键词: 图像去雾, 编解码网络, 扩张残差, 特征增强, 通道注意力

Abstract:

Image dehazing is a typical ill-posed problem.Dehazing network architectures commonly use encoder-decoder networks consisting of encoders, decoders, and feature converters that connect the two.The quality of dehazing images generated by existing dehazing algorithms is generally low, with problems such as incomplete local detail dehazing, color distortion, or the introduction of noise.Dehazing algorithms based on an encoder-decoder network do not fully utilize small-scale features when designing feature converters, and only use the corresponding layer coding features in the decoding stage; therefore, this paper proposes a dehazing algorithm based on high-frequency and low-frequency feature enhancement.In the feature transformation stage, an expanded residual component is designed and formed into a context aggregation network, which makes full use of the low-resolution features of the large receptive field, extracts the remote correlation of feature maps, and enhances low-frequency small-scale features.Furthermore, a multi-level feature reuse network based on channel attention is designed to reuse shallow high-frequency features while deeply fusing decoded and reconstructed features to enhance the recovery of visual perceptual features.In the coding stage, a Visual Feature Perception Module(VFPM) is constructed to enhance the rich high-frequency visual features of the shallow layer by leveraging the advantages of residual blocks in local modeling.Experimental results show that compared with AOD-Net, PFF-Net, and other dehazing algorithms, the Peak Signal-to-Noise Ratio(PSNR) and Structural Similarity(SSIM) of the proposed algorithm have significant advantages.The PSNR and SSIM of the proposed algorithm are 0.77 dB and 0.000 7, and 0.40 dB and 0.037 1 higher than those of the other algorithms, respectively, on an indoor synthetic dataset, SOTS, and outdoor real dataset, Dense-Haze.

Key words: image dehazing, encoder-decoder network, dilated residual, feature enhancement, channel attention