利用深度學習之YOLOv3偵測機車車牌

Locomotive License Plate Detection via Deep Learning YOLOv3

蔡俊明(Chun-Ming Tsai) ； 林宗易(Zong-Yi Lin)

《國教新知》 65卷4期 (2018/12) Pp. 9-20

https://doi.org/10.6701/TEEJ.201812_65(4).0002

摘要

在現在的交通工具中，機車還是最便利，最不會塞車的交通工具，然而，機車失竊常常發生，機車車主要找回失竊的機車，實在不容易。另外，一些犯罪，常偷竊機車來犯案，警察要緝凶，也常以車追人，然而機車那麼多，警察要調閱各路口監視器，慢慢看慢慢找，才有可能找到，甚至找不到。為了幫助機車失主和警察，能快速找到失竊機車和作案兇手，若能有一套智慧型機車車號偵測和辨識系統，就可以幫助機車失主和警察，尋回其愛車和逮補兇手。然而機車車牌的偵測對智慧型機車車號偵測和辨識系統最為重要，車牌偵測不到，就不用辨識車牌了。因為在真實的環境中，機車車牌會出現在白天、夜晚、模糊、旋轉、強光、甚至陰暗中，這些車牌利用傳統車牌偵測方法來偵測，其效果有限。為了能偵測上述真實環境中的機車車牌，本文利用深度學習YOLOv3技術來訓練和偵測真實環境中的機車車牌。我們的做法是先標記機車車牌708張影像作為車牌偵測訓練資料，另外288張機車影像作為測試資料。實驗結果顯示，機車車牌偵測準確率可達97.72%。

關鍵字

深度學習 ； YOLOv3 ； 機車車牌偵測

並列摘要

In the current transportation, the locomotive is still the most convenient transportation. However, the locomotive stealing often occurs, and it is not easy to find the stolen locomotive. In addition, some crimes often steal locomotive to commit crimes. The police must seize these thief. They often find the thief by following the locomotive. However, there are so many locomotives. The police should see many videos at various intersections and look for the stolen locomotive to find them. However, it is very slowly and even cannot find them. In order to help the locomotive owner and the police to find the stolen locomotive and the thief quickly, an intelligence locomotive license plate detection and recognition system should be proposed. However, the detection of locomotive license plates is the most important for this system. If the license plate is not detected, there is no need to identify the license plate. Because in real environment, locomotive license plates will appear in daylight, night, blur, rotation, glare, and even darkness. These license plates are detected by traditional license plate detection methods, and their effects are limited. In order to detect the locomotive license plates in the real environment mentioned above, this article uses the deep learning YOLOv3 technology to train and detect the locomotive license plates in the real environment. Our approach is to label the 708 images of the locomotive license plates as the training samples and uses another 288 locomotive images to be tested. The experimental results show that the detection accuracy of locomotive license plate can reach 97.72%.

並列關鍵字

deep learning ； yolov3 ； license plate detection

參考文獻

Arth, C.,Limberger, F.,Bischof, H.(2007).Real-Time License Plate Recognition on an Embedded DSP-Platform.Conference on Computer Vision and Pattern Recognition.(Conference on Computer Vision and Pattern Recognition).:

Duan, T.,Du, T.,Phuoc, T.,Hoang, N.(2005).Building an Automatic Vehicle License-Plate Recognition System.Conference: International. Conference in Computer Science - RIVF'05.(Conference: International. Conference in Computer Science - RIVF'05).:

Guo, J.,Liu, Y.(2008).License Plate Localization and Character Segmentation With Feedback Self-Learning and Hybrid Binarization Techniques.IEEE Trans. Veh. Technol..57(3),1417-1424.

Nikolaos, C.,Ioannis.,Loumos, V.,Kayafas, E.(2006).A License Plate-Recognition Algorithm for Intelligent Transportation System Applications.IEEE Transactions on Intelligent Transportation Systems.7(3),377-392.

Sovani, M.,Vo, D.,Challa, S.,Palaniswami, M.(2015).A Feedback based method for License plate image binarization.2015 9th International Conference on Signal Processing and Communication Systems (ICSPCS).(2015 9th International Conference on Signal Processing and Communication Systems (ICSPCS)).