臺灣博碩士論文加值系統

機器學習是人工智慧的一個重要領域，自從以來機器學習在算法、理論以及應用都有廣大的成功案例，而機器學習領域當中有個名為「深度學習」的課題漸漸受到學術界的重視，至今已成為人工智慧和互聯網大數據的熱題。深度學習通過建立類似人腦中使用神經元的方式，對輸入的資料提取從底層到高層的特徵，例如現在的CNN、RNN、DNN等深度學習架構都是廣泛被全世界使用。舉例來說，無人駕駛汽車會使用深度學習來處理影像，或者區分路上的物件或者行人，這些應用在我們生活中都扮演極度重要的角色。然而這些功能都與影像處理息息相關，例如車輛計數系統，在遇到夜晚時段辨識出來的圖片，因路燈或者車燈的嚴重光害的關係導致辨識效果不良，可能會導致收集到的資訊不正確，為了解決因光害造成的辨識問題，而提出本論文。
本論文主要目的是解決因過度曝光的圖像而影響無法準確定位車輛的問題，提出一種先將圖像經過亮度補償再進行二值化處理及圖像除霧(Dehazing)的預處理，最後再利用YOLOv4深度學習演算法去做訓練和測試。主要處理步驟如下：首先對圖像進行預處理，主要目的對圖像中不同亮度區域進行亮度補償，並且利用除霧功能將圖片模糊部分變得更清晰，來提升除光害後生成圖像的品質。把預處理過後的圖像採取7：3的資料分配，7成資料為訓練資料，而3成資料為測試資料。再將訓練資料給YOLOv4去做學習，並用測試資料來取得目標偵測的評估指標。本方法主要探討將過度曝光的車輛影像做預處理，將計算mAP (Mean Average Precision)值呈現出經過預處理過後和無經過預處理的差異性。

Machine learning is an important area of artificial intelligence (AI). Machine learning has been successfully applied many applications. A topic called "deep learning" in the machine learning has gradually attracted the attention of the academia. It has become a hot research in the fields of AI and internet big data. Deep learning extracts low-level to high-level features from input data by building methods which is similar to the way the human brain using neurons. Currently, deep learning architectures such as CNN, RNN, and DNN are widely used worldwide. For example, self-driving cars use deep learning to process images, or to distinguish between objects or pedestrians on the road, all of which play an extremely important role in our lives. However, these functions are closely related to image processing. For example, when a vehicle counting system encounters an image recognized at night, the recognition effect is poor due to light pollution from street lamps or vehicle lights, which may lead to incorrect information collected. In order to solve the identification problem caused by light pollution, this thesis proposes to design a feasible method to overcome the related image problems caused by overexposure.
The main purpose of this thesis is to solve the problem that the vehicles cannot be positioned accurately due to light pollution. The proposed method consists of the pre-processing of the brightness compensation, image binarization and dehazing, and finally uses YOLOv4 to achieve this system. The main steps are as follows: First, preprocessing the image, this step purpose is to compensate the brightness of different brightness areas in the image. And using the dehazing process, the blurred part of the image makes more clearly thereby improved the clarity of the image. The resulting image quality can be improved after removing light pollution. Afterward, the experimental data are distributed in the ratio of 7:3, where 70% of the data is used as training data and 30% is used as test data. Then, the training data is handed over to YOLOv4 for learning, and the test data is used to verify the system performance. This method will evaluate mAP value to analyze the difference between preprocessed and unprocessed ones applied to car localization with high light exposure.

摘要...i
Abstract...iii
誌謝...v
目錄...vi
表目錄...viii
圖目錄...ix
符號說明...xi
第一章 緒論...1
1.1簡介...1
1.2研究動機與目的...2
1.3論文架構...3
第二章 文獻探討及基礎理論...4
2.1 YOLO神經網路(You only look once, YOLO)...4
2.1.1網路設計...5
2.1.2 池化層(Pooling layer)...6
2.2 DeepMultiBox...7
2.2.1 Training Objective...8
2.3相關文獻...10
第三章 系統架構...15
3.1 Gamma 校正(Gamma correction, GC)...16
3.2 影像二值化處理...19
3.3 除霧處理[7] (Dehazing process)...22
3.4 YOLOv4 (You Only Look Once)...34
第四章 實驗結果與討論...38
實驗環境平台...38
4.2實驗資料庫...38
4.3實驗參數設定...39
YOLOv4...39
Gamma (γ)效正...40
4.4實驗分析及結果...40
4.5討論...44
Max_batches 選擇...44
Gamma (γ)選擇...48
第五章 結論與未來展望...52
參考文獻...53
Extended Abstract...56

[1]Joseph Redmon, Santosh Divvala, and Ross Girshick, and Ali Farhadi, “You Only Look Once:Unified, Real-Time Object Detection,” in Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016, pp.779-788.
[2]Dumitru Erhan, Christian Szegedy, Alexander Toshev, and Dragomir Anguelov Google, “Scalable object detection using deep neural networks,” in Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014, pp.2147-2154.
[3]Meng Wei and Meng Tia, “Unmanned aerial vehicle classification and detection based on deep transfer learning,” in International Conference on Intelligent Computing and Human-Computer Interaction (ICHCI), 2020, pp.6617-6621.
[4]Hangzhou Zhang and Yi Wan, “A novel enhancement algorithm for night image by adjusting the background,” in 2016 8th International Conference on Intelligent Human-Machine Systems and Cybernetics (IHMSC). 2016, pp.304-307.
[5]Yan Miao, Fu Liu, Tao Hou, Lu Liu, and Yun Liu, “A nighttime vehicle detection method based on YOLO v3,” 2020 Chinese Automation Congress (CAC), pp. 6617-6621.
[6]Linrunjia Liu, Cindy Cappelle, and Yassine Ruichek, “Day and night place recognition based on low-quality night-time images,” in the IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), 2020, pp.1-6.
[7]Zhang Jing, Yang Cao, and Zengfu Wan, “Nighttime haze removal based on a new imaging model,” in Proc. of IEEE International Conference on Image Processing (ICIP), 2014, pp.4557-4561.
[8]Thathupara Subramanyan Kavya, Erdenetuya Tsogtbaatar, Young-Min Jang, and Sang-Bock Cho, “Night-time vehicle detection based on brake/Tail light color,” 2018 International SoC Design Conference (ISOCC), 2018, pp. 206-207.
[9]Muhamad Yani, S, Si., M.T. Budhi Irawan and S.T., M.T. Casi Setiningsih, “Application of transfer learning using convolutional neural network method for early detection of terry’s nail,” Journal of Physics: Conference Series, vol. 1201, 2019. https://doi.org/10.1088/1742-6596/1201/1/012052
[10]Gaofeng Meng, Ying Wang, Jiangyong Duan, Shiming Xiang, and Chunhong Pan, “Efficient image dehazing with boundary constraint and contextual regularization,” in Proc. of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2013, pp.617-624.
[11]Castleman and Kenneth R, Digital image processing. Prentice Hall Press, 1996.
[12]曾炫皓，空間變異之適應性參數的gamma校正運用於高曝光與低曝光影像之研究。2013 國立勤益科技大學碩士論文。
[13]S. G. Narasimhan and S. K. Nayar, “Contrast restoration of weather degraded images,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 6, pp.713-724, 2003.
[14]Raanan Fattal, “Single image dehazing.” ACM Transactions on Graphics, vol. 27, no. 3, pp. 1-9, 2008.
[15]Kaiming He, Sun Jian, and Tang Xiaoou, “Single image haze removal using dark channel prior,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 12, pp. 2341-2353, 2010.
[16]Robby T. Tan, “Visibility in bad weather from a single image,” in Proc. of the IEEE Conference on Computer Vision and Pattern Recognition 2008, pp. 1-8.
[17]Louis Kratz and Ko Nishino, “Factorizing scene albedo and depth from a single foggy image,” in Proc. of the IEEE 12th International Conference on Computer Vision, 2009, pp.1701-1708.
[18]Bochkovskiy Alexey, Chien-Yao Wang, and Hong-Yuan Mark Liao, “Yolov4: Optimal speed and accuracy of object detection,” arXiv:2004.10934v1, 2004.
[19]Zhuliang Yao, Yue Cao, Shuxin Zheng, Gao Huang, and Stephen Lin, “Cross-iteration batch normalization,” https://arxiv.org/abs/2002.05712v3, 2020.
[20]Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia, “Path aggregation network for instance segmentation.” in Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 8759–8768.
[21]Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon, “CBAM: Convolutional block attention module,” in Proc. of the European Conference on Computer Vision (ECCV), 2018, pp. 3-19.
[22]Jiang Da Bai, “深入淺出 Yolo 系列之 Yolov3 & Yolov4 & Yolov5核心基礎知識完整講解,” retrieval May, 2020 from https://zhuanlan.zhihu.com/p/143747206.
[23]Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia, “Path Aggregation Network for Instance Segmentation,” https://doi.org/10.48550/arXiv.1803.01534.
[24]顏大立，基於YOLOv4之即時旗語辨識研究，2021 中原大學碩士論文。
[25]t74_Database, https://www.kaggle.com/datasets/vivia0525/t74-database