卷積類神經演算法於影像風格轉換之研究

A study on the convolutional neural algorithm of image style transfer

鮑奕諠(Yi-Hsuan Pao)

指導教授 ： 林慧珍

淡江大學/工學院/資訊工程學系資訊網路與多媒體碩士班/碩士(2016年)

https://doi.org/10.6846/TKU.2016.01110

摘要

近年來深度類神經網路學習在各領域的發展都有很好的成效，在影像分類上已有顯著的進步，應用在影像的藝術風格轉換，也有很好的成果。L. A. Gatys et al. [6]所提出的利用學習好的CNN (Convolutional Neural Network)架構 VGG [13]來做影像風格轉換，也有令人為之驚豔的結果。本研究將針對L. A. Gatys et al. 所提出的影像風格轉換法，其中遇到的問題，做探討並提出可能的改進或解決方法。在作影像的藝術風格轉換訓練階段，隨著網路層數的增加，其運算時間也急速膨脹，尤其在倒傳遞修正過程中，勢必會遇到許多不同的問題(風格影像和內容影像比例參數調整、修正量調整，加速運算，⋯)。本研究主要探討這些問題的解決方法，包含如何有效地選擇能量函數權重值、如何化簡偏微分乘積鏈、如何加速修正量的計算，經實驗證明運算時間有明顯的改善，我們所提出正規化權重調整的方法，也顯現令人滿意的效果。

關鍵字

最大池化 ； 倒傳遞 ； 風格轉換 ； 過度擬合 ； 深度卷積類神經網路 ； 全卷積網路 ； 感知域 ； 卷積核 ； 合併卷積核

並列摘要

Recently, deep convolutional neural networks have resulted in noticeable improvements in image classification and are used to transfer artistic style of images. L. A. Gatys et al. [6] proposed the use of a learned CNN (Convolutional Neural Network) architecture VGG [16] to transfer image style, but problems occur during the back propagation process because there is a heavy computational load. This paper solves these problems, including the simplification of the computation of chains of derivatives, accelerating the computation of adjustments, and efficiently choosing weights for different energy functions. The experimental results show that the proposed solutions improve the computational efficiency and render the adjustment of weights for energy functions easier.

並列關鍵字

max-pooling ； back-propagation ； style transfer ； over-fitting ； deep convolutional neural networks (DCNN) ； fully convolutional networks ； receptive field ； kernel ； merge kernel

參考文獻

[1] O. Abdel-Hamid, A. Mohamed, and H. Jiang, "Convolutional neural networks for speech recognition," IEEE/ACM Trans. Audio, Speech and Lang.Proc., vol. 22, pp. 1533-1545, Oct., 2014.

[2] S. Bell and K. Bala, "Learning visual similarity for product design with convolutional neural networks," ACM Trans. Graph., vol. 34, pp. 98:1-98:10, Aug., 2015.

[3] Y. Bengio, "Learning deep architectures for AI," Foundations and Trends in Machine Learning, vol. 2, pp. 1-127, 2009.

[4] A. Dosovitskiy, J. T. Springenberg, and T. Brox, "Learning to generate chairs with convolutional neural networks," CoRR, abs/1411.5928, 2014. Available: http://arxiv.org/abs/1411.5928.

[5] K. Fukushima, "Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position," Biological Cybernetics, vol. 36, pp. 193-202, 1980.