圖信號之採樣與快速的局部卷積神經網路之物體識別應用

Downsampling of Graph Signals and Object Detection Application Using Fast Region-based Convolutional Networks

洪培軒(Pei-Hsuan Hung)

指導教授 ： 貝蘇章

國立臺灣大學/電機資訊學院/電信工程學研究所/碩士(2016年)

https://doi.org/10.6342/NTU201600471

摘要

本論文由兩部分組成，在第一部分中，我們研究了圖信號(Graph Signals)下採樣方法，圖信號處理是一個新興的信號處理領域讓我們可以處理分析結構不規則信號，時至今日變得越來越重要。而這些圖形信號的操作也成為了最近的許多研究的主題，尤其是對於諸如移動，調變和採樣的基本信號的操作。然而，在應用中的圖的尺寸可以非常大，所以導致儲存或分析這些資料成為了一個重大的技術挑戰。要在圖表更有效地壓縮這些數據，我們提出了一個預過濾分類器(Pre-filtering Classifer)，其會考慮信號在圖上的分佈，可以有選擇性地進行信號的採樣。相比於其它方法，如著色法(Color-based)和基於拓撲方法(Topology-based)，我們提出的方法可以得到較高的信噪比。而且我們的方法可以很有效率的在短時間之內得到一定的壓縮效果。

本論文的第二部分會談到有關如何使用快速局部卷積神經網路(Fast R-CNN)開發一些物體偵測的應用，從運算環境的設置包括GPU的建置和可平行運算的快速局部卷積神經網路的平台的建置，到如何進行該演算法的學習以及測試。近年來通過使用基於局部卷積神經網絡，物體檢測的正確性有著大幅的進展，而快速局部卷積神經網路演算法可幫助我們即時的得到成果縱使是使用非常深層的網絡架構時。為了更加理解這種高效能方法，我們提出了它的一些應用。另外類似的機器學習技術也可以用在圖信號處理中。

關鍵字

圖信號處理 ； 訊號壓縮 ； 預過濾分類器 ； 卷積神經網路 ； 物體偵測

並列摘要

This thesis consists of two sections. In the first section, we study the downsampling methods for graph signals. Graph Signal Processing is an emerging field of signal processing for us to analysis irregular structure signals and becomes more and more significant in these days. The operations on these datasets as graph signals have been subjects to many recent studies, especially for basic signal operations such as shifting, modulating, and down-sampling. However, the sizes of the graphs in the applications can be very large and lead a lot of computational and technical challenges for the purpose of storage or analysis. To compress these datasets on graphs more effectively, we propose a pre-filtering classifier can selectively downsample signals and also consider the distribution of the signals on graphs. As compared to the other methods, such as color-based methods and topology-based methods, our proposed method can achieve better performance in terms of higher SNR. Moreover, our method can be processed efficiently and efficacy in terms of shorter computing-time and fewer vertices in use during compression.

The second section of this thesis talks about how to use Fast Regions with Convolutional Neural Network (Fast R-CNN) to develop some object detection applications from the building of the environment including the setup of GPU and the platform of parallel computing to the process of training and testing in fast R-CNN algorithm. By using region-based convolutional neural networks, the correctness of object detection has a large progress in recent years, and fast R-CNN algorithm helps us to achieve near real-time rates when using very deep networks. To realize this efficient and powerful method more, some applications based on it are also proposed. Further, a machine learning technique is also applied to graph signal processing.

並列關鍵字

Graph signal processing ； Signal compression ； pre-filtering classifier ； nonvolutional neural network ； object detection

參考文獻

[2] G. Shen and A. Ortega, “Optimized distributed 2D transforms for irregularly sampled sensor network grids using wavelet lifting,” in Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), pp. 2513–2516, 2008.

[3] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, “The emerging field of signal processing on graphs: extending high-dimensional data analysis to networks and other irregular domains,” IEEE Signal Processing Magazine, vol. 30, no. 3, pp. 83–98, 2013.

[4] M. Tsitsvero and S. Barbarossa. “On the degrees of freedom of signals on graphs.” in Proceedings of the Signal Processing Conference (EUSIPCO), pp. 1531-1535, 2015.

[5] B. Girault. “Stationary graph signals using an isometric graph translation.” in Proceedings of the Signal Processing Conference (EUSIPCO), pp. 1521-1525, 2015.

[6] G. Shen and A. Ortega, “Transform-based distributed data gathering,” IEEE Transactions on Signal Processing, vol. 58, no. 7, pp. 3802–3815, 2010.