郁車牌: 行車中的自動車牌辨識

Abstract

本篇論文提出了一個專為移動車輛設計的車牌辨識 (License Plate Recogniotion) 系統解決方案，旨在實現在車輛行駛期間實時進行車牌辨識。該系統整合了工業電腦、紅外攝影機和電源供應器，即將開發一套名為「AI (Artificial Intelligence) 人工智慧機車路邊停車計費系統」。本論文通過改進的深度學習網絡架構進行車牌辨識，以應對車輛行駛過程中拍攝到的不同角度、光線、移動模糊等因素，從而提高辨識準確性。為了兼顧計算資源有限和電力消耗的考慮，本論文選擇了較輕量的網絡架構。該系統的辨識過程分為兩個階段：首先，通過一個卷積神經網絡 (Convolutional Neural Network，CNN) 架構對街道場景進行車牌偵測，然後使用另一個卷積神經網絡對偵測到的車牌影像進行光學字元辨識，包括大寫字母(A-Z)和數字(0-9)。該方法在白天以 20 公里/小時高速行駛的影片測試實現了95.8%的準確率、97.1%的召回率及96.45%的 F1-score，並在 10 公里/小時的實際行駛實驗中達到了98.06%的準確率。

This thesis proposes a solution for real-time License Plate Recognition (LPR) on moving vehicles. Our future system, named the "AI Motorcycle Parking Fee Collection System," integrates Industrial Personal Computers (IPC), infrared cameras, and power supplies. YuLPR employs an advanced deep learning network architecture to address challenges such as varying angles, lighting conditions, and motion blur encountered during vehicle movement, thereby enhancing recognition accuracy. To balance computational efficiency and power consumption, a lightweight network architecture is adopted. The recognition process involves two stages: initial license plate detection using a Convolutional Neural Network (CNN) on street scenes, followed by Optical Character Recognition (OCR) using another CNN on the detected license plate images, encompassing uppercase letters (A-Z) and digits (0-9). YuLPR achieves a precision of 95.8%, a recall of 97.1%, and a F1-score of 96.45% in video tests conducted at a high speed of 20 kilometers per hour (km/h) during daytime, and reaches an accuracy of 98.06% in actual riding experiments conducted at a speed of 10 km/h.