具批量正規化之卷積去雜訊自編碼器性能分析

Performance Analysis of Convolutional Denoising Autoencoder with BatchNormalization

胡智超(Jhih-Chao Hu)

指導教授 ： 陳文雄

國立暨南國際大學/科技學院/電機工程學系/碩士(2020年)

https://doi.org/10.6837/ncnu202000291

摘要

雜訊在影像辨識當中造成很大的問題，如何將有雜訊的影像還原出較好的品質，將是本論文的重點。本論文提出了二種方法對於現有的卷積去雜訊自編碼器架構[36]進行改進，第一種方法將卷積去雜訊自編碼器每層卷積層加入批量正規化，實驗使用高斯雜訊、遮罩雜訊以及各種雜訊強度透過優化器做分析，探討加入批量正規化對於影像還原的效益。第二種方法將原有的卷積去雜訊自編碼器的架構進行改良，改以堆疊卷積之架構，再以每層卷積層後加入批量正規化，三種架構進行比較分析，最後經由CNN做辨識。本論文在MNIST、MHD、Cifar-10、UTKFace等資料集進行實驗，在MNIST資料集中，二種方法皆較原始架構所還原的影像品質要好，經CNN辨識結果測試準確率有略微提升，在MHD、Cifar-10、UTKFace等資料集中，CNN辨識測試準確率皆有明顯提升，還原影像的品質得到提升。

關鍵字

卷積神經網路 ； 卷積去雜訊自編碼器 ； 批量正規化 ； 堆疊卷積去雜訊自編碼器

並列摘要

Noise causes great problems in image recognition. How to restore images with noise to a better quality will be the focus of this thesis. This thesis proposes two methods to improve the existing convolutional denoising autoencoder architecture [36]. The first method adds each convolutional layer of the convolutional denoising autoencoder to batch normalization. The experiment uses gaussian noise, mask noise, and various noise intensities are analyzed through the optimizer to discuss the benefits of adding batch normalization to image restoration. The second method improves the architecture of the original convolution denoising autoencoder, changeing to a stacked convolution architecture, and adds batch normalization after each convolution layer. The three architectures are compared and analyzed. Finally, CNN does the identification. This thesis conducts experiments on data sets such as MNIST, MHD, Cifar-10 and UTKFace, In the MNIST data set, both methods are better than the image quality restored by the original architecture. The test accuracy of the CNN recognition results is slightly improved. In the MHD,Cifar-10 and UTKFace datasets, the test accuracy of the CNN recognition has been significantly improved, and the quality of restored images has been improved.

並列關鍵字

Convolutional Neural Network ； Convolutional Denoising Autoencoder ； Batch Normalization ； Stack Convolutional Denoising Autoencoder

參考文獻

[1] P. Vincent, H. Larochelle, Y. Bengio and P.-A. Manzagol,“Extracting and Composing Robust Features with Denoising Autoencoders,” Proceedings of the Twenty-Fifth International Conf. (ICML 2008) , June 5-9, 2008.

[2] A. Pretorius, S. Kroon and H. Kamper, “Learning Dynamics of Linear Denoising Autoencoders,” 35th International Conf. on Machine Learning (ICML) 2018, 2018.

[3] X. Lu , Y. Tsao , S. Matsuda and C. Hori, “Speech Enhancement Based on Deep Denoising Autoencoder,” INTERSPEECH 2013, 2013.

[4] R. E. Zezario, J.-W. Huang, X. Lu, Y. Tsao, H.-T. Hwang and H.-M. Wang, “Deep Denoising Autoencoder Based Post Filtering for Speech Enhancement,” 2018 Asia-Pacific Signal and Information Processing Association Annual Summit and Conf. (APSIPA ASC), 2018.

[5] J. Jiao, L. Bao, Y. Wei, S. He, H. Shi, R. Lau, T. S. Huang, “Laplacian Denoising Autoencoder,” ICLR Conf. Blind Submission, 2019.