| 摘要: | 由於科技進步，人類對於交通工具的需求已是不可或缺，對於台灣來說，車輛的數量更幾乎等同於台灣的總人口數，再加上台灣地小人多的關係，如何有效率的進行車輛停放管理更加重要，而車牌辨識系統的優劣即是處理這個問題的關鍵。目前停車場出入口多半為定位式車牌辨識系統，而路邊停車收費則必須使用移動式車牌辨識系統，移動式車牌系統需要在更多種環境下進行辨識，特別是歪斜的狀況是最常發生的，相較於定位式牌辨識系統更具挑戰性。受惠於近年來深度學習的發展，提高影像處理技術上限，同時快速地進行多張歪斜車牌的辨識已是輕而易舉，因此本研究的目標是將深度學習應用到車牌辨識系統，並將其架設到停車收費管理員的行動裝置上且達到辨識歪斜角度在0-45度的車牌有99%的準確率與單張圖像0.5秒以下的辨識速度，為此，本研究使用歪斜車牌辨識模型IWPODNet與字元辨識模型YOLOv5 並先在電腦端分別以1490張車牌資料集與5560張字元資料集將兩個模型作訓練，其中IWPODNet和YOLOv5的模型解析度選用480X368與160X160，並在308張測試資料IWPODNet得到0.8的Recall，YOLOv5得到了Precision為0.982、Recall為0.973以及F-Score為0.976的良好成果。而辨識速度則分別為單張圖像0.798秒及0.045秒，在APP階段，以模型解析度大小分析準確度與速度的權衡後，最後決定兩個模型分別以288X216與160X160的模型解析度轉換成Tensorflow-Lite模型以便在行動裝置上操作，再架設到合作公司提供的RS35 Android行動電腦上，並建立車牌辨識App，並且使用正則表達式作為字元結果呈現的篩選，最後使用App進行實際拍攝304張200萬畫素的照片，拍攝狀況是以停車收費員視角進行拍攝，並針對距離3公尺內1台汽車或1到3台機車為目標。最後得到90.7%的Precision以及單張圖像最快0.7秒的辨識速度。從此結果也看出IWPODNet較適合在電腦端模擬且利於發揮其轉正的優勢，而YOLOv5不論在的電腦端及APP端都有快速且優良的辨識能力。;With the advancement of technology, the demand for transportation has become indispensable for humans. In the case of Taiwan, the number of vehicles is almost equivalent to the total population, considering the small size of the country and the high population density. Therefore, efficient vehicle parking management is crucial. The performance of license plate recognition systems plays a key role in addressing this issue. Currently, most parking lot entrances and exits use fixed-position license plate recognition systems, while mobile license plate recognition systems are required for roadside parking fee collection. Mobile systems need to perform recognition in a wider range of environments, particularly skewed angles, making them more challenging compared to fixed-position systems. Thanks to the development of deep learning in recent years, improving the upper limit of image processing technology and quickly recognizing multiple license plates has become effortless. Therefore, the goal of this study is to apply deep learning to license plate recognition systems and deploy them on the mobile devices of parking fee administrators, achieving a recognition accuracy of 99% for license plates with skew angles ranging from 0 to 45 degrees and a recognition speed of less than 0.5 seconds per image. To accomplish this, the study utilizes the slanted license plate recognition model IWPODNet and the character recognition model YOLOv5. Both models are trained using separate datasets consisting of 1490 license plate images and 5560 character images on a computer. The resolutions selected for IWPODNet and YOLOv5 models are 480x368 and 160x160. On the 308 test datas, IWPODNet achieved good results with recall of 0.8 and YOLOv5 achieved good results with precision of 0.982, and recall of 0.973. The recognition speeds are 0.798 seconds and 0.045 seconds per image, respectively. In the app phase, after analyzing the trade-off between accuracy and speed based on model resolution size, the two models are finally converted into TensorFlow Lite models with resolutions of 288x216 and 160x160, respectively, for operation on mobile devices.We create a license plate recognition app and use regular expressions as filters for presenting character results. Our models are deployed on the RS35 Android mobile computer provided by the collaborating company to develop a license plate recognition app. Finally, the app is tested by capturing 304 photos with a resolution of 2 million pixels, from the perspective of a parking fee collector, targeting 1 car or 1 to 3 motorcycles within a distance of 3 meters. The results show a precision of 90.7% and the fastest recognition speed of 0.7 seconds per image. From these results, it can be observed that IWPODNet is more suitable for simulation on a computer and benefits from its advantage in handling slanted license plates. On the other hand, YOLOv5 demonstrates fast and excellent recognition capabilities both on the computer and in the app. |
