计算机系统应用  2018, Vol. 27 Issue (9): 170-175 PDF

1. 中国科学院大学, 北京 100049;
2. 中国科学院 沈阳计算技术研究所, 沈阳 110168;
3. 沈阳市环境检测中心站, 沈阳 110000;
4. 沈阳市第二十七中学, 沈阳 110011

Image Reconstruction Algorithm Based on Deep Convolution Neural Network
YU Bo1,2, FANG Ye-Quan1,2, LIU Min3, DONG Jun-Tao4
1. University of Chinese Academy of Sciences, Beijing 100049, China;
2. Shenyang Institute of Computing Technology, Chinese Academy of Sciences, Shenyang 110168, China;
3. Shenyang Environmental Monitoring Center Station, Shenyang 110000, China;
4. Shenyang Twenty-Seventh Middle School, Shenyang 110011, China
Abstract: In the process of video or image transmission, there may be random error, sudden error, packet loss, and so on, which will also have a serious impact on the decoded image data. This paper presents an image reconstruction algorithm based on depth learning: an unsupervised image reconstruction neural network model based on image background prediction to generate fuzzy region content. In order to reconstruct a vivid image, a neural network model not only needs to understand the content of the image, but also to reconstruct the missing part of a reasonable assumption. The loss function includes standard pixel level reconstruction loss and counterwork loss. When training the convolution neural network model, the loss function can better deal with the structure details in the image and produce clearer results. Through experiments, we can find that the neural network model of depth convolution designed in this study has better effect in image reconstruction than the algorithm based on sample interpolation.
Key words: image reconstruction     depth learning     neural network     loss function     adversarial neural network

1 卷积神经网络模型总体设计

 图 1 神经网络结构示意图

1.1 卷积提取图像特征

1.2 反卷积重建图像

1.3 重建图像损失

2 神经网络模型各个模块的原理

2.1 特征提取

 图 2 可视化神经网络模型

 $W = \frac{1}{9}\left( {\begin{array}{*{20}{c}} 1&1&1&1 \\ 1&1&1&1 \\ 1&1&1&1 \\ 1&1&1&1 \end{array}} \right)$ (1)
 图 3 图像卷积示意图

2.2 图像重建

2.3 损失函数

 ${l_\gamma} = d\left( {G(Z) - Z} \right) = {\left\| {G(Z) - Z} \right\|^2}$ (2)

 图 4 对抗神经网络示意图

 $\max \log_{10} (D(G(z)))$ (3)

 ${l_a} = \mathop {\max }\limits_D (\log_{10} (D(x)) + \log_{10} (1 - D(G(z))))$ (4)

 $\ell {\rm{ = }}{\lambda _\gamma}{l_\gamma} + {\lambda _\alpha}{l_\alpha}$ (5)

3 卷积神经网络模型训练与实验

 $PSNR = 10 \times {\log _{10}}\left( {\frac{{{{\left( {{2^n} - 1} \right)}^2}}}{{MSE}}} \right)$ (6)

4 结论与展望

 图 5 四种算法的图像重建实验结果

 [1] Mikolov T, Sutskever I, Chen K, et al. Distributed representations of words and phrases and their compositionality. Advances in Neural Information Processing Systems 26. 2013. 2–3. [2] LeCun Y, Boser B, Denker JS, et al. Backpropagation applied to handwritten zip code recognition. Neural Computation, 1989, 1(4): 541-551. DOI:10.1162/neco.1989.1.4.541 [3] Dosovitskiy A, Springenberg JT, Brox T. Learning to generate chairs with convolutional neural networks. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA, USA. IEEE. 2015. [doi: 10.1109/CVPR.2015.7298761] [4] Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation. IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA, USA. IEEE. 2015. [5] Zeiler MD, Fergus R. Visualizing and understanding convolutional networks. In: Fleet D, Pajdla T, Schiele B, et al., eds. Computer Vision–ECCV 2014. Lecture Notes in Computer Science, vol 8693. Springer, Cham. 2014. 818–833. [doi: 10.1007/978-3-319-10590-1_53] [6] Goodfellow I, Pouget-Abadie J, Mirza M, et al. Generative adversarial nets. arXiv, 2014: 1406.2661. [7] Krizhevsky A, Sutskever I, Hinton GE. ImageNet classiﬁcation with deep convolutional neural networks. Communications of the ACM, 2017, 60(6): 84-90. DOI:10.1145/3065386 [8] Hays J, Efros AA. Scene completion using millions of photographs. ACM SIGGRAPH, 2007. DOI:10.1145/1239451.1239455 [9] Dong C, Loy CC, He KM, et a1. Learning a deep convolutional network for image super-resolution. In: Fleet D, Pajdla T, Schiele B, et al., eds. Computer Vision–ECCV 2014. Lecture Notes in Computer Science, vol 8692. Springer, Cham. 2014. 184–199. [doi: 10.1007/978-3-319-10593-2_13] [10] Martin D, Fowlkes C, Tal D, et al. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. Proceedings of the Eighth IEEE International Conference on Computer Vision. Vancouver, BC, Canada. 2001. 416–423. [doi: 10.1109/ICCV.2001.937655]