基於卷積神經網路之車牌辨識系統

Abstract

車牌的偵測與辨識用途廣泛，例如：停車場收費系統、道路監視器以及測速照相機等。大多數的解決辦法為使用傳統影像分析技術來進行處理，相關的研究方法以及改進的方案已經發展多年，辨識的成功率也都很高，但都建立在兩個重要的前提下：一、車牌必須清晰，不能受到太多光影干擾，也不能有大量的髒污附著在上面；台灣的車牌除了老舊污損，也常貼滿各種貼紙標誌，影響辨識結果。二、車牌不能夠過於歪斜，因此拍攝視角往往是同一角度及位置，否則在車牌偵測與字元切割上會受到嚴重影響，進而造成辨識不易。為了解決這兩點，本論文採用一模型 “YOLO”(You Only Look Once)，該模型是一種機器學習(machine learning)中基於深度學習(deep learning)的卷積神經網路(convolutional neural network，CNN)，利用卷積層(convolution layer)來擷取目標物的特徵進而達到辨識的效果。整個辨識過程共使用三組網路，第一組從影像中偵測車牌，第二組用以偵測車牌上的字元，最後將偵測成功的字元送入第三組網路進行字元辨識。實驗結果顯示，利用卷積神經網路無須滿足上述兩點在傳統影像處理方法上重要的前提下，也能夠將車牌及其字元辨識成功。

Car plate detection and recognition is widely used, such as parking fees system, road monitor and speeding camera. Most of the solutions use traditional image analysis techniques to process, related research methods and improved solutions get high recognition accuracy which have been developed for many years. But all of them have two important prerequisite. First, car plate must be clear, it can’t be affected by light and have a lot off dirt attached to the above. Second, car plate can’t be too skewed so that the angle and position of shooting view are often the same. Otherwise the car plate detection and character segmentation will be seriously affected and then cause recognition difficulty. In order to solve these two points, our research use a model “YOLO” (You Only Look Once). This model is a convolutional neural network based on the deep learning in machine learning. It use convolution layers to get the features of object and then achieve the effect of recognition. We use a total of three models in the research. Detecting car plates from the image in the first model, and then use the second model to detect characters on the car plates which have been detected by the first model. Finally, input characters which are detected successfully to the third model to do character recognition. The results show that our research can recognize car plate and its characters successful by using convolutional neural network without satisfy two important prerequisite based on traditional image process methods.