《计算机应用》唯一官方网站 ›› 2024, Vol. 44 ›› Issue (2): 572-579.DOI: 10.11772/j.issn.1001-9081.2023020123

• 多媒体计算与计算机仿真 • 上一篇    

融合监督注意力模块和跨阶段特征融合的图像修复改进网络

黄巧玲1, 郑伯川1,2(), 丁梓成2, 吴泽东2   

  1. 1.西华师范大学 数学与信息学院,四川 南充 637009
    2.西华师范大学 计算机学院,四川 南充 637009
  • 收稿日期:2023-02-15 修回日期:2023-05-08 接受日期:2023-05-11 发布日期:2023-08-14 出版日期:2024-02-10
  • 通讯作者: 郑伯川
  • 作者简介:黄巧玲(1998—),女,四川绵阳人,硕士研究生,主要研究方向:机器学习、深度学习、图像修复
    丁梓成(1999—),男,四川巴中人,硕士研究生,主要研究方向:机器学习、深度学习、图像修复
    吴泽东(2000—),男,四川眉山人,硕士研究生,主要研究方向:机器学习、深度学习、目标检测。
  • 基金资助:
    国家自然科学基金资助项目(62176217);西华师范大学科研创新团队项目(KCXTD2022?3)

Improved image inpainting network incorporating supervised attention module and cross-stage feature fusion

Qiaoling HUANG1, Bochuan ZHENG1,2(), Zicheng DING2, Zedong WU2   

  1. 1.School of Mathematics & Information,China West Normal University,Nanchong Sichuan 637009,China
    2.School of Computer Science,China West Normal University,Nanchong Sichuan 637009,China
  • Received:2023-02-15 Revised:2023-05-08 Accepted:2023-05-11 Online:2023-08-14 Published:2024-02-10
  • Contact: Bochuan ZHENG
  • About author:HUANG Qiaoling, born in 1998, M. S. candidate. Her research interests include machine learning, deep learning, image inpainting.
    DING Zicheng, born in 1999, M. S. candidate. His research interests include machine learning, deep learning, image inpainting.
    WU Zedong, born in 2000, M. S. candidate. His research interests include machine learning, deep learning, object detection.
  • Supported by:
    National Natural Science Foundation of China(62176217);Scientific Research Innovation Team Project of China West Normal University(KCXTD2022-3)

摘要:

非规则缺失区域的图像修复技术用途广泛但具有挑战性。针对现有修复方法对高分辨率图像可能会产生伪影、扭曲结构和模糊纹理的问题,提出一种融合监督注意力模块(SAM)和跨阶段特征融合(CSFF)的图像修复改进网络(Gconv_CS)。在Gconv的两阶段网络模型上,引入了SAM与CSFF模块。SAM通过提供真实图像监督信号,监督上阶段输出特征,确保传入下阶段特征信息的有效性。CSFF将上阶段编码器-解码器的特征融合后送入下阶段的编码器,以弥补上阶段修复中特征信息的损失。实验结果表明,在缺失区域占比为1%~10%时,相较于基线模型Gconv,Gconv_CS在CelebA-HQ数据集上峰值信噪比(PSNR)和结构相似性指数(SSIM)分别提高了1.5%和0.5%,Fréchet起始距离(FID)和L1损失分别降低了21.8%、14.8%;在Place2数据集上,前2个指标分别提高了26.7%和0.8%,后2个指标分别降低了7.9%、37.9%。将Gconv_CS用于去除大熊猫面部遮挡物时,取得了较好的修复视觉效果。

关键词: 图像修复, 两阶段网络, 跨阶段特征融合, 监督注意力模块, 门控卷积

Abstract:

Image inpainting techniques for non-regular missing regions are versatile but challenging. To address the problem that existing inpainting methods may produce artifacts, distorted structures, and blurred textures for high-resolution images, an improved image inpainting network, named Gconv_CS(Gated convolution based CSFF and SAM) incorporating Supervised Attention Module (SAM) and Cross-Stage Feature Fusion (CSFF) was proposed. In Gconv_CS, the SAM and CSFF were introduced to Cconv, a two-stage network model with gated convolution. SAM ensured the effectiveness of the incoming feature information to the next stage by providing a real image to supervise the output features of the previous stage. CSFF fused the features from the encoder-decoder of the previous stage and fed them to the encoder of the next stage to compensate for the loss of feature information in the previous stage. The experimental results show that, at a percentage of missing regions of 1% to 10%, compared with the baseline model Gconv, on CelebA-HQ dataset, Gconv_CS improved the Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity index (SSIM) by 1.5% and 0.5% respectively, reduced Fréchet Inception Distance (FID) and L1 loss by 21.8% and 14.8% respectively; on Place2 dataset, the first two indicators increased by 26.7% and 0.8% respectively, and the latter two indicators decreased by 7.9% and 37.9% respectively. A good restoration effect was achieved when Gconv_CS was used to remove masks from a giant panda’s face.

Key words: image inpainting, two-stage network, Cross-Stage Feature Fusion (CSFF), Supervised Attention Module (SAM), gated convolution

中图分类号: