| 本文已被：浏览 5463次 下载 8717次 |
码上扫一扫！ |
|
|
| 基于小波域的深度残差网络图像超分辨率算法 |
|
段立娟1,2,3, 武春丽1,2,3, 恩擎1,2,3, 乔元华4, 张韵东5, 陈军成1
|
|
1.北京工业大学 信息学部, 北京 100124;2.可信计算北京市重点实验室, 北京 100124;3.信息安全等级保护关键技术国家工程实验室, 北京 100124;4.北京工业大学 应用数理学院, 北京 100124;5.数字多媒体芯片技术国家重点实验室(北京中星微电子有限公司), 北京 100191
|
|
| 摘要: |
| 单幅图像超分辨率（SISR）是指从一张低分辨率图像重建高分辨率图像.传统的神经网络方法通常在图像的空间域进行超分辨率重构，但这些方法常在重构过程中忽略重要的细节.鉴于小波变换能够将图像内容的"粗略"和"细节"特征进行分离，提出一种基于小波域的深度残差网络（DRWSR）.不同于其他传统的卷积神经网络直接推导高分辨率图像（HR），该方法采用多阶段学习策略，首先推理出高分辨率图像对应的小波系数，然后重建超分辨率图像（SR）.为了获取更多的信息，该方法采用一种残差嵌套残差的灵活可扩展的深度神经网络.此外，提出的神经网络模型采用结合图像空域与小波域的损失函数进行优化求解.所提出的方法在Set5、Set14、BSD100、Urban100等数据集上进行实验，实验结果表明，该方法的视觉效果和峰值信噪比（PSNR）均优于相关的图像超分辨率方法. |
| 关键词: 单幅图像超分辨率 小波变换 卷积神经网络 残差块 |
| DOI：10.13328/j.cnki.jos.005663 |
| 分类号: |
| 基金项目:国家重点研发计划（2017YFC0803705）；国家自然科学基金（61572004，61771026）；北京市自然基金委-市教委联合资助项目（KZ201910005008）；青海省创新平台建设专项（2016-ZJ-Y04） |
|
| Deep Residual Network in Wavelet Domain for Image Super-resolution |
|
DUAN Li-Juan1,2,3, WU Chun-Li1,2,3, EN Qing1,2,3, QIAO Yuan-Hua4, ZHANG Yun-Dong5, CHEN Jun-Cheng1
|
|
1.Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China;2.Beijing Key Laboratory of Trusted Computing, Beijing 100124, China;3.National Engineering Laboratory for Key Technologies of Information Security Level Protection, Beijing 100124, China;4.College of Applied Science, Beijing University of Technology, Beijing 100124, China;5.State Key Laboratory of Digital Multimedia Chip Technology(Vimicro Corp.), Beijing 100191, China
|
| Abstract: |
| Single Image Super Resolution (SISR) refers to the reconstruction of high resolution images from a low resolution image. Traditional neural network methods typically perform super-resolution reconstruction in the spatial domain of an image, but these methods often ignore important details in the reconstruction process. In view of the fact that wavelet transform can separate the "rough" and "detail" features of image content, this study proposes a wavelet-based deep residual network (DRWSR). Different from other traditional convolutional neural networks, the high-resolution image (HR) is directly derived. This method uses a multi-stage learning strategy to first infer the wavelet coefficients corresponding to the high-resolution image and then reconstruct the super-resolution image (SR). In order to obtain more information, the method uses a flexible and scalable deep neural network with residual nested residuals. In addition, the proposed neural network model is optimized by combining the loss function of image space and wavelet domain. The proposed method is carried out on Set5, Set14, BSD100, Urban100, and other datasets. The experimental results show that the proposed visual effect and peak signal-to-noise ratio (PNSR) are better than the related image super-resolution method. |
| Key words: single image super-resolution wavelet transform convolutional neural network residual block |
