臺灣博碩士論文加值系統

在過去深度圖像的去雜訊方法通常是針對雜訊種類套用特定模組來去除，而近年來在一般圖像的去雜訊上越來越多使用深度學習的方法，因此在此篇論文中我們嘗試使用深度學習，利用深度卷積神經網路針對深度圖像進行去雜訊，同時研究在單一訓練模組下可去除的雜訊種類以及程度多寡，我們用blender來產生隨機3D場景的深度圖像當作訓練以及測試用的資料集，我們在這項研究中使用的雜訊種類有白高斯雜訊、椒鹽雜訊、斑點雜訊、萊利雜訊和對數常態雜訊，在雜訊等級 15 的情況下，白高斯雜訊、椒鹽雜訊、斑點雜訊去噪最高可達到 PSNR 平均值個別等於 41.30 dB、37.19 dB、35.93dB，萊利雜訊去噪是特殊例子，去雜訊後的 PSNR 值在15左右但是在視覺上是可以接受的，對數常態雜訊則是失敗的案例。

In the past, the methods of denoising depth images were usually by using a specific module for the specific type of noise. In recent years, the deep learning technique is employed to denoise general images. We propose to use the deep learning method, which is based on the Convolutional Neural Networks (CNN) to denoise the depth images. We also research that under a single training module how many types of noise can be reduced and how wide the noise level range can be handled. To generate the training sets and testing sets, we use Blender to produce depth images from random 3D scenes. The types of noise we employed in this study are the Additive white Gaussian noise (AWGN), the Salt & Pepper Noise, the Speckle Noise, the Rayleigh Noise, and the Lognormal Noise. When the noise level is set to 15, the AWGN, the Salt & Pepper Noise, and the Speckle Noise denoising can achieve the highest PSNR mean of 41.30 dB, 37.19 dB. and 35.93 dB, respectively. The Riley noise denoising is a special case, the PSNR mean is 15 dB after denoising but the denoised images are visually acceptable. The Lognormal Noise denoising is a failed case.

論文摘要……………………...…………………….………………………….…I
Abstract……………………...……………..……….………………………….…II
Contents…………………...……………..………...………………………….…III
List of Figures……….………………..………...………...………………….…IV
List of Tables……….……….………..………....………...………...….…….…VI
Chapter 1. Introduction…………………………………………………………...1
Chapter 2. Related Work…………..……………………...……………………...3
2.1 Deep Learning………………………………………………………....3
2.2 Convolutional Neural Network (CNN)…….…………………………4
2.3 Category of Noise…….….……………………………………………5
Chapter 3. Employed CNN Model…………..………………………………….15
3.1 FFDNet model…………….….….….….….…….……...…………....15
3.2 Network Architecture………………………………………………....15
3.3 Noise Level Map and Denoising on Sub-images……………………16
Chapter 4. Experiment…………………………………………………………..19
4.1 Dataset Generation………….….….….….….….……………………..19
4.2 Modifying and Training the Model……………………………………22
4.3 Experiments on Noise Removal with Default Model………………….22
4.4 Experiments on Noise Removal with Modified Models………………29
Chapter 5. Conclusions and Future work…………………...…..…..…………..40
Reference…………………………….…………………...……………………..41

[1] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising by sparse 3-D transform-domain collaborative filtering,” IEEE Transactions on Image Processing, vol. 16, no. 8, pp. 2080–2095, 2007.
[2] K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising,” IEEE Transactions on Image Processing, vol. 26, no. 7, pp. 3142–3155, July 2017.
[3] K. Zhang, W. Zuo, and L. Zhang, “FFDNet: Toward a fast and flexible solution for CNN based image denoising,” IEEE Trans. Image Process., vol. 27, no. 9, pp. 4608–4622, 2018.
[4] K. Lai, L. Bo, X. Ren, and D. Fox, “A large-scale hierarchical multi-view RGB-D object dataset,” in Proc. IEEE Int. Conf. Robot. Autom., 2011, pp. 1817–1824.
[5] Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A review and new perspectives,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 35, no. 8, pp. 1798–1828, Aug. 2013.
[6] J. Schmidhuber, “Deep learning in neural networks: An overview,” Neural Networks, vol. 61, pp. 85–117, 2015.
[7] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, 2015.
[8] Talking about the principle and application of Deep Learning, “http://www.cc.ntu.edu.tw/chinese/epaper/0038/20160920_3805.html”, Referenced on May 24 th, 2019.
[9] A. Krizhevsky, I. Sutskever, and G. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” Proc. Neural Information and Processing Systems, 2012.
[10] Convolutional neural network,
“https://en.wikipedia.org/wiki/Convolutional_neural_network”, Referenced on May 25 th, 2019.
[11] Additive white Gaussian noise,
“https://en.wikipedia.org/wiki/Additive_white_Gaussian_noise”, Referenced on May 25 th, 2019.
[12] Image Restoration,
“http://www.nhu.edu.tw/~CSIE/ycliaw/DIP/05_Image_Restoration.pdf”, Referenced on May 25 th, 2019.
[13] Salt-and-pepper noise, “https://en.wikipedia.org/wiki/Salt-and-pepper_noise”, Referenced on May 25 th, 2019.
[14] Forouzanfar, M., Abrishami-Moghaddam, H. “Ultrasound Speckle Reduction in the Complex Wavelet Domain” Principles of Waveform Diversity and Design. SciTech Publishing an imprint of the IET, pp. 558–577, 2010.
[15] Rayleigh distribution,
“https://en.wikipedia.org/wiki/Rayleigh_distribution#cite_note-PP-2”, Referenced on May 25 th, 2019.
[16] D. Kundu and M. Z. Raqab, “Generalized Rayleigh distribution: Different methods of estimations,” in Computational Statistics Data Analysis, vol. 49, pp. 189–200, 2005.
[17] Lognormal Distribution,
“https://www.sciencedirect.com/topics/engineering/lognormal-distribution”, Referenced on May 25 th, 2019.
[18] Log-normal distribution,
“https://en.wikipedia.org/wiki/Log-normal_distribution” Referenced on May 25 th, 2019.
[19] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105.
[20] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in International Conference on Machine Learning, 2015, pp. 448–456.
[21] Blender, “https://en.wikipedia.org/wiki/Blender_(software)”, Referenced on May 24 th, 2019
[22] DepthMap_dataset, “https://github.com/LouisFoucard/DepthMap_dataset”, Referenced on May 24 th, 2019.
[23] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in International Conference for Learning Representations, 2015.
[24] K. D. Bonin and M. A. Kadar-Kallen, “Simple diffuser for production of laser speckle,” Applied Optics, vol. 28, no. 24, pp. 5293–5297, 1989.
[25] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in International Conference on Machine Learning, pp. 448–456, 2015.