本論文提出了設計輕量級類神經網絡的簡單方法。深度學習領域有許多先進的神經網路，在常見的數據集（例如:ImageNet和Cifar-10，Cifar-100）有非常良好的準確度，除了執行時間短之外模型體積也不大。然而這些現有的先進類神經網路結構非常的複雜。本論文提出了一種構建輕量級類神經網路的簡便方法，亦即，使用分離式卷積輕量化模型體積，同時縮減層數尺度不變之等校層數以加速運算，另外，提出一種學習模型切換機制，在不同天候下的路面偵測效能，能切換至最適合的對應神經網路模型，獲得最佳的路面偵測性能。從kitti dataset 與台北地區路面資料庫的實驗結果中，證明本論文所提出的輕量化類神經網路可模型切換之可行駛路面偵測，具有不錯的偵測效能，適合未來自駕車進行可行駛區域偵測之參考。

This study proposes a set of simple methods to design a light-weight neural network. A deep learning domain has many state of the art neural networks so it is highly accurate for commonly used dataset such as ImageNet and Cifar-10, Cifar-100 and allows a rapid execution time and a small model. However, these state of the art neural networks are very complicated. This paper uses a VGG-16 model to reduce the size of the model and the inference time, but maintain accuracy. The semantic segmentation performance for the proposed method is compared to that for the VGG-16 model. The same full convolutional network (FCN) semantic segmentation algorithm is used to compare the two models for the same semantic segmentation task. This study proposes an easier method to construct a light-weight neural network. We also using the model switch algorithm to apply in different weather condition. Many semantic segmentation dataset for road surface just has only one weather condition. But in the real world, the weather has different kind of situation that will influence the result of the road detection. So we train models separately in different weather condition. Then we use the model switch algorithm to fit in different weather. Our contribution are making simple process to realize the light weight neural network and model switch algorithm.

摘要 i
ABSTRACT ii
誌謝 iv
目錄 v
圖目錄 ix
表目錄 xiv
第一章 緒論 1
1.1前言 1
1.2研究動機與目的 3
1.3本論文貢獻 4
1.4 本論文架構 4
第二章 文獻探討 5
2.1基於影像處理之路面偵測 6
2.1.1基於影像結構之路面偵測 6
2.1.2基於影像像素之路面偵測 6
2.2基於機器學習之路面偵測 7
2.2.1支援向量機 7
2.2.2類神經網路 8
2.2.3類神經網路分類器 10
2.2.4物件偵測 11
2.2.5語意分割 12
2.2.6即時性語意分割 17
2.2.7光達與影像融合之技術 18
2.2.8多深度學習模型之應用 19
2.3總結 20
第三章 語意分割類神經網路輕量化與模型切換技術 21
3.1語意分割架構介紹 21
3.1.1全卷積網路架構 21
3.1.2基底類神經網路介紹 23
3.2類神經網路輕量化技巧介紹 24
3.2.1分離式卷積 24
3.2.2縮減層數尺度不變之等校層數 26
3.2.3縮減層數尺度不變之替換池化 27
3.3多模型切換技術 29
3.3.1多模型切換之概念 29
3.3.2分類器之介紹 30
3.3.3模型切換演算法詳述 32
3.4總結 33
第四章 實驗結果 34
4.1實驗設備與軟體平台 34
4.2實驗流程 34
4.3平台架設 37
4.3.1 Ubuntu18.04系統之架設 37
4.3.2 Windows10實驗平台架設 38
4.4 Kitti dataset與相關實驗 39
4.4.1 Kitti dataset與實驗結果 39
4.4.2 Kitti dataset在台灣場景之實驗結果 42
4.5 台灣道路環境資料製作 52
4.5.1 台灣道路之晴天環境資料製作 53
4.5.2 台灣道路之陰天環境資料製作 56
4.5.3 台灣道路之白天雨天環境資料製作 58
4.5.4 台灣道路之夜間雨天環境資料製作 63
4.6 VGG16-FCN在台灣場景之實驗 64
4.6.1 台灣道路之晴天環境訓練與測試 64
4.6.2 台灣道路之陰天環境訓練與測試 67
4.6.3 台灣道路之雨天環境訓練與測試 69
4.6.4 台灣道路之夜間雨天環境訓練與測試 72
4.7 類神經網路改良與方法改進 74
4.7.1 類神經網路輕量化 74
4.7.2 類神經網路輕量化:增大訓練批次量 82
4.7.3 方法改進:模型切換演算法(Model switch algorithm) 82
4.8 輕量化網路與模型切換之執行結果 88
4.8.1 LWF-VGG在Kitti dataset上的實驗結果 88
4.8.2 LWF-VGG在台灣晴天狀態之實驗結果 92
4.8.3 LWF-VGG在台灣陰天狀態之實驗結果 94
4.8.4 LWF-VGG在台灣白天雨天狀態之實驗結果 96
4.8.5 LWF-VGG在台灣夜間雨天狀態之實驗結果 98
4.8.6 模型切之實驗結果 101
4.9 總結 103
第五章 結論與未來展望 104
5.1結論 104
5.2未來展望 105
參考文獻 106

[1]Forrest N. Iandola, Song Han, Matthew W. Moskewicz, Khalid Ashraf, William J. Dally, Kurt Keutzer, SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size , 4 Nov 2016.
[2]Amir Gholami, Kiseok Kwon, Bichen Wu, Zizheng Tai, Xiangyu Yue, Peter Jin, Sicheng Zhao, Kurt Keutzer, SqueezeNext: Hardware-Aware Neural Network Design, 27 Aug 2018.
[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, 17 Apr 2017.
[4]Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen, MobileNetV2: Inverted Residuals and Linear Bottlenecks, 21 Mar 2019.
[5]Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, Hartwig Adam, Searching for MobileNetV3, 20 Nov 2019.
[6] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, Jian Sun, ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices, 7 Dec 2017.
[7]Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, Jian Sun, ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design, 30 Jul 2018.
[8]Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, Deep Residual Learning for Image Recognition, 10 Dec 2015.
[9]Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, Zbigniew Wojna, Rethinking the Inception Architecture for Computer Vision, 2 Dec 2015.
[10]Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He, Aggregated Residual Transformations for Deep Neural Networks, 11 Apr 2017.
[11]Bowen Baker, Otkrist Gupta, Nikhil Naik, Ramesh Raskar, Designing Neural Network Architectures using Reinforcement Learning, 22 Mar 2017.
[12]Barret Zoph, Vijay Vasudevan, Jonathon Shlens, Quoc V. Le, Learning Transferable Architectures for Scalable Image Recognition, 11 Apr 2018.
[13]Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, Kevin Murphy, Progressive Neural Architecture Search, 26 Jul 2018.
[14]Esteban Real, Alok Aggarwal, Yanping Huang, Quoc V Le, Regularized Evolution for Image Classifier Architecture Search, 16 Feb 2019
[15]Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, Quoc V. Le, MnasNet: Platform-Aware Neural Architecture Search for Mobile, 29 May 2019.
[16]Mingxing Tan, Quoc V. Le, EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks, 23 Nov 2019.
[17]Karen Simonyan, Andrew Zisserman, Very Deep Convolutional Networks for Large-Scale Image Recognition, 10 Apr 2015.\
[18]Joseph Redmon, Ali Farhadi, YOLOv3: An Incremental Improvement, 8 Apr 2018.
[19]Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg, SSD: Single Shot MultiBox Detector, 29 Dec 2016.
[20]Ross Girshick, Jeff Donahue, Trevor Darrell, Jitendra Malik, Rich feature hierarchies for accurate object detection and semantic segmentation, 22 Oct 2014.
[21]Ross Girshick, Fast R-CNN, 27 Sep 2015.
[22]Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, Nong Sang, BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation, 2 Aug 2018.
[23]Changqian Yu, Changxin Gao, Jingbo Wang, Gang Yu, Chunhua Shen, Nong Sang, BiSeNet V2: Bilateral Network with Guided Aggregation for Real-time Semantic Segmentation, 5 Apr 2020.
[24]Adam Paszke, Abhishek Chaurasia, Sangpil Kim, Eugenio Culurciello, ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation, 7 Jun 2016.
[25]Zhe Chen, Jing Zhang, Dacheng Tao, Progressive LiDAR Adaptation for Road Detection, 2 Apr 2019.
[26]Irem Ulku, Erdem Akagunduz, A Survey on Deep Learning-based Architectures for Semantic Segmentation on 2D images, 14 May 2020.
[27]Jonathan Long, Evan Shelhamer, Trevor Darrell, Fully Convolutional Networks for Semantic Segmentation, 8 Mar 2015.
[28]Tatiana Tommasi, Francesco Orabona, Barbara Caputo, Learning Categories From Few Examples With Multi Model Knowledge Transfer, 29 May 2014.
[29]Kamran Kowsari, Mojtaba Heidarysafa, Donald E. Brown, Kiana Jafari Meimandi, Laura E. Barnes, RMDL: Random Multimodel Deep Learning for Classification, 31 May 2018.
[30]Samuel L. Smith, Pieter-Jan Kindermans, Chris Ying, Quoc V. Le, Don't Decay the Learning Rate, Increase the Batch Size, 24 Feb 2018.
[31]Ping Chao, Chao-Yang Kao, Yu-Shan Ruan, Chien-Hsiang Huang, Youn-Long Lin, HarDNet: A Low Memory Traffic Network, 3 Sep 2019
[32]資料網站，Ubuntu安裝，https://andy6804tw.github.io/2019/05/12/ubuntu-install/
[33]資料網站，顯示卡安裝，https://linuxhint.com/ubuntu_nvidia_ppa/
[34]資料網站，Cuda安裝，https://developer.nvidia.com/cuda-toolkit-archive/
[35]資料網站，cudaNN安裝，https://www.itread01.com/content/1546932988.html
[36]資料網站，路面分割VGG16-FCN程式碼，https://github.com/penny4860/Kitti-road-semantic-segmentation
[37]資料網站，anaconda安裝，https://www.anaconda.com/products/individual
[38]資料網站，pip指令介紹，https://www.maxlist.xyz/2019/07/13/pip-install-python/
[39]資料網站，conda指令介紹，https://jennaweng0621.pixnet.net/blog/post/403957928-anaconda%E6%8C%87%E4%BB%A4%E6%87%89%E7%94%A8%E7%B8%BD%E6%95%B4%E7%90%86
[40]資料網站，git指令介紹, https://tedsieblog.wordpress.com/2017/01/19/basic-git-commands/
[41]資料網站，rufus開機資料碟燒入程式，https://rufus.ie/zh_TW.html
[42]Olaf Ronneberger, Philipp Fischer, Thomas Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation, 18 May 2015
[43]Vijay Badrinarayanan, Alex Kendall, Roberto Cipolla, SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation, 10 Oct 2016
[44]Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition, 23 Apr 2015
[45]Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, Jiaya Jia, Pyramid Scene Parsing Network, 27 Apr 2017
[46]Fisher Yu, Vladlen Koltun, Multi-Scale Context Aggregation by Dilated Convolutions, 30 Apr 2016
[47]Huikai Wu, Junge Zhang, Kaiqi Huang, Kongming Liang, Yizhou Yu, FastFCN: Rethinking Dilated Convolution in the Backbone for Semantic Segmentation, 28 Mar 2019
[48]Dao-Wen Tang, Chih-Ming Hsu, A Vision-based Around View Road Detection , 16 July 2016
[49]資料網站，路面標示照片，https://www.picfair.com/pics/05316754-keep-clear-road-marking
[50]資料網站，影片轉圖片之程式，https://www.pkstep.com/archives/37880
[51]資料網站，像素標記之工具程式，https://www.mathworks.com/help/vision/ug/get-started-with-the-image-labeler.html