臺灣博碩士論文加值系統

本文採用YOLOv3及RetinaNet之深度學習演算法對汽機車車牌進行辨識。首先進行圖像資料採集，資料源自於交通大學智慧校園校門監視系統影像和個人行車紀錄器影像截圖。為了模擬真現實移動狀況，本文嘗試加入大量失焦圖像以及手動加入雜訊圖像於訓練資料集中，測試兩模型是否可以保有原本清楚車牌辨識能力並同時兼顧對於失焦與模糊車牌的偵測。從只使用定點單一車牌所訓練出之YOLOv3模型，其精確率和召回率為85%與96%；而RetinaNet模型之精確率和召回率都可達到99%，展現出比較優秀的結果。但在混和行車紀錄器與失焦圖進行訓練後，YOLOv3的精確率和召回率可上升到90%和99%；RetinaNet相較於第一次數值，則下降到93%與95%，與YOLOv3最多只有3%差距，但YOLOv3辨識所需時間只為RetinaNet的三分之一。同時實驗兩模型在白天與夜晚的狀態下之偵測率，本文歸納出RetinaNet較適用於停車場等定點偵測，YOLOv3比較適合動態偵測。

This thesis used deep learning algorithms of YOLOv3 and RetinaNet to make the license plate recognition. The data came from two sources, one from gate monitoring system of the National Chiao Tung University and the other from a personal video recorder carried by a motorcycle rider. In order to simulate the out-of-focus scenario of a video recorder in real situation, a number of the out-of-focus images and the images with noise were added in the training data to test whether these two algorithms can still maintain the original recognition of license plate and identify the out-of-focus or fuzzy license plate simultaneously. The results of YOLOv3 are that the recall and precision rates are 86% and 96%, respectively, in the recorder situation of a fixed position for training data. As to RetinaNet, it achieves 99% in both of values. When adding noise and other conditions, such as the different backgrounds or lighting, to images in training data, the precision rate is reduced from 99% to 93%, and recall rate is from 99% to 95%. On the other hand, the precision rate is increased from 86% is to 90%, and the recall rate is from 96% up to 99% in YOLOv3’s model. Although two models have almost the same performance, the running time of YOLOv3 is faster than that of RetinaNet by triple. The detection rates of the two models in the daytime and nighttime states were tested. It is found that RetinaNet is more suitable for fixed-point detection, such as in a parking lot, whereas YOLOv3 is more suitable in a moving detection.

摘要 I
ABSTRACT II
誌謝 III
目錄 IV
表目錄 V
圖目錄 VI
第一章 緒論 1
1.1 研究動機與目的 1
1.2 文獻回顧 2
1.3 研究方法與架構 5
第二章 資料收集與處理 7
2.1資料來源 7
2.2 資料類型 12
2.3.資料處理與標註 17
第三章 模型方法 19
3.1神經網路(NEURAL NETWORK) 19
3.2 RETINANET 20
3.3 YOLOV3(YOU ONLY LOOK ONCE VERSION 3, YOLOV3) [16] 26
3.4優化器(OPTIMIZER) 29
3.4.1梯度下降法(Gradient descent) 30
3.4.2小批量梯度下降法(Mini-batch Gradient descent) 30
3.4.3動量梯度下降法(Momentum Gradient descent) [22] 31
3.4.4 Adagrad [23] 31
3.4.5 Adam [24] 32
3.5評估方法(EVALUATION) 33
第四章 結果與討論 34
第五章 結論 53
參考文獻 55

[1] A. M. Al-Ghaili, S. Mashohor, A. R. Ramli, and A. J. I. t. o. v. t. Ismail, "Vertical-edge-based car-license-plate detection method," vol. 62, no. 1, pp. 26-38, 2012.
[2] C.-T. Hsieh, Y.-S. Juan, and K.-M. Hung, "Multiple license plate detection for complex background," in 19th International Conference on Advanced Information Networking and Applications (AINA'05) Volume 1 (AINA papers), 2005, vol. 2, pp. 389-392: IEEE.
[3] L. J. C. S. Dlagnekov and S. D. Engineering Department, "License plate detection using adaboost," 2004.
[4] Y. Freund, R. Schapire, and N. J. J.-J. S. F. A. I. Abe, "A short introduction to boosting," vol. 14, no. 771-780, p. 1612, 1999.
[5] C. N. E. Anagnostopoulos, I. E. Anagnostopoulos, V. Loumos, and E. J. I. T. o. I. t. s. Kayafas, "A license plate-recognition algorithm for intelligent transportation system applications," vol. 7, no. 3, pp. 377-392, 2006.
[6] Y. LeCun, L. Bottou, Y. Bengio, and P. J. P. o. t. I. Haffner, "Gradient-based learning applied to document recognition," vol. 86, no. 11, pp. 2278-2324, 1998.
[7] O. Bulan, V. Kozitsky, P. Ramesh, and M. J. I. T. o. I. T. S. Shreve, "Segmentation-and annotation-free license plate recognition with deep localization and failure identification," vol. 18, no. 9, pp. 2351-2363, 2017.
[8] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet classification with deep convolutional neural networks," in Advances in neural information processing systems, 2012, pp. 1097-1105.
[9] R. Polishetty, M. Roopaei, and P. Rad, "A next-generation secure cloud-based deep learning license plate recognition for smart cities," in 2016 15th IEEE International Conference on Machine Learning and Applications (ICMLA), 2016, pp. 286-293: IEEE.
[10] Z. Selmi, M. B. Halima, and A. M. Alimi, "Deep learning system for automatic license plate detection and recognition," in 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR), 2017, vol. 1, pp. 1132-1138: IEEE.
[11] K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770-778.
[12] G.-S. Hsu, A. Ambikapathi, S.-L. Chung, and C.-P. Su, "Robust license plate detection in the wild," in 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS), 2017, pp. 1-6: IEEE.
[13] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, "You only look once: Unified, real-time object detection," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 779-788.
[14] J. Redmon and A. Farhadi, "YOLO9000: better, faster, stronger," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 7263-7271.
[15] X. Hou, M. Fu, X. Wu, Z. Huang, and S. Sun, "Vehicle License Plate Recognition System Based on Deep Learning Deployed to PYNQ," in 2018 18th International Symposium on Communications and Information Technologies (ISCIT), 2018, pp. 79-84: IEEE.
[16] J. Redmon and A. J. a. p. a. Farhadi, "Yolov3: An incremental improvement," 2018.
[17] S. Abdullah, M. M. Hasan, and S. M. S. Islam, "YOLO-Based Three-Stage Network for Bangla License Plate Recognition in Dhaka Metropolitan City," in 2018 International Conference on Bangla Speech and Language Processing (ICBSLP), 2018, pp. 1-6: IEEE.
[18] Lin-Tzu-Ta, in labelImg, ed.
[19] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, "Focal loss for dense object detection," in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2980-2988.
[20] S. Ren, K. He, R. Girshick, and J. Sun, "Faster r-cnn: Towards real-time object detection with region proposal networks," in Advances in neural information processing systems, 2015, pp. 91-99.
[21] T.-Y. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie, "Feature pyramid networks for object detection," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 2117-2125.
[22] N. J. N. n. Qian, "On the momentum term in gradient descent learning algorithms," vol. 12, no. 1, pp. 145-151, 1999.
[23] J. Duchi, E. Hazan, and Y. J. J. o. M. L. R. Singer, "Adaptive subgradient methods for online learning and stochastic optimization," vol. 12, no. Jul, pp. 2121-2159, 2011.
[24] D. P. Kingma and J. J. a. p. a. Ba, "Adam: A method for stochastic optimization," 2014.