臺灣博碩士論文加值系統

近年來深度學習快速地發展，其中卷積神經網路(Convolutional Neural Network)，常被用於不同偵測及辨識的問題當中，在影像處理方面效果更是令人讚嘆。透過卷積神經網路特徵擷取的能力，能夠偵測不同環境下的路牌，並辨識路牌資訊。
本論文是以深度學習作為基礎，對國道上的交通路牌進行偵測與辨識，讓駕駛者可以透過此系統得知前方路牌資訊。偵測方面，使用深度可分離卷積結構(Depthwise Separable Convolution)，與單階段的物件偵測模型，在複雜的背景中偵測路牌。辨識方面，利用前饋式類神經網路(FeedForward Neural Network，FFNN)來辨識路牌資訊。實驗結果表明路牌偵測模型的精確率(Precision)以及召回率(Recall)有不錯的表現且不同的特徵擷取架構會影響移動式裝置上的測試時間而路牌辨識模型的準確率(Accuracy)達到99%的表現。

Deep learning is developing rapidly in recent years, which Convolutional Neural Network(CNN) often used for different detection and recognition problems, especially for image process. The capability of feature extraction by Convolutional Neural Network can detect panels in different environment and recognize information of panels. In this paper, we propose an application for the detection and recognition of highway traffic panels. Drivers can realize information of panels by our application. In the detection stage, we use one-stage object detection model and Depthwise Separable Convolution to detect panels in complex background. In the recognition stage, we use feedforward neural network to recognize information of panels. The experimental results show that panel detection model precision and recall are good and different feature extraction layer will effect inference time on mobile devices, panel recognition model achieve 99% accuracy.

誌謝 i
摘要 ii
Abstract iii
目錄 iv
圖目錄 vii
表目錄 viii
第一章 緒論 1
1.1 研究動機 1
1.2 研究目的 1
1.3 研究方法 1
1.4 章節流程 2
第二章 相關研究 3
2.1前饋式神經網路(FEED FORWARD NEURAL NETWORK) 3
2.1.1激活函數(Activation Function) 4
2.1.2反向傳導演算法(Backpropagation Algorithm) 4
2.2卷積神經網路(CONVOLUTIONAL NEURAL NETWORKS) 5
2.2.1卷積層(Convolutional Layer) 5
2.2.2池化層(Pooling Layer) 6
2.3 SSD(SINGLE SHOT MULTIBOX DETECTOR) 7
2.3.1多尺度特徵圖偵測 7
2.3.2 預設框與長寬比 8
2.3.3匹配策略 9
2.4 深度可分離卷積(DEPTHWISE SEPARABLE CONVOLUTION) 11
2.4.1 Depthwise convolution 12
2.4.2 Pointwise convolution 12
第三章 研究方法 14
3.1 系統架構 14
3.2 路牌偵測模型 15
3.2.1多尺度特徵圖偵測 17
3.2.2預設框與長寬比 18
3.2.3 非極大值抑制(Non-Maximum Suppression，NMS) 19
3.3 路牌辨識模型 20
第四章 實驗結果 21
4.1 實驗環境 21
4.2資料集 21
4.3 TENSORFLOW模型轉換 22
4.4 模型評估 23
4.5訓練 23
4.5.1資料增強(Data Augmentation) 23
4.5.2 Hard negative mining 24
4.6實驗結果 25
4.6.1路牌偵測模型 26
4.6.2路牌辨識模型 27
4.6.3時間測試 27
第五章 結論 29
參考文獻 30

[1] Simon Haykin. “Feedforward Neural Networks: An Introduction,” Neural Networks: A Comprehensive Foundation (3rd Edition) , 2007.
[2] Nair, V. and Hinton, G.E. “Rectified Linear Units Improve Restricted Boltzmann Machines. Proceedings of the 27th International Conference on Machine Learning,” Haifa, pp.807-814, 2010.
[3] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, Hartwig Adam. “MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,” arXiv:1704.04861, 2017.
[4] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg. “SSD: Single Shot MultiBox Detector,” arXiv: 1512.02325, 2016.
[5] Ross Girshick, Jeff Donahue, Trevor Darrell, Jitendra Malik. “Rich feature hierarchies for accurate object detection and semantic segmentation,” arXiv: 1311.2524, 2014.
[6] Ross Girshick. “Fast R-CNN,” arXiv: 1504.08083, 2015.
[7] Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun. “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,” arXiv: 1506.01497, 2015.
[8] J.R.R. Uijlings, K.E.A. van de Sande, T. Gevers, A.W.M. Smeulders. “Selective search for object recognition,” International Journal of Computer Vision, Volume 104, Issue 2, pp. 154–171, 2013.
[9] Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi. “You Only Look Once: Unified, Real-Time Object Detection,” arXiv: 1506.02640, 2015.
[10] Karen Simonyan, Andrew Zisserman. “Very Deep Convolutional Networks for Large-Scale Image Recognition.” arXiv:1409.1556, 2015.
[11] Sergey Ioffe, and Christian Szegedy. “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,” arXiv:1502.03167, 2015.
[12] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen. “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” arXiv: 1801.04381, 2018.