號碼布之影像辨識研究

Abstract

本論文研究影像中號碼布的辨識方法，主要分為號碼布定位及字元辨識兩個部分。 號碼布定位利用了邊緣偵測(Edge Detection)、形態學(Morphology)中的侵蝕(Erosion)與膨脹(Dilation)再搭配機器學習(Machine learning)中的支持向量機(Support Vector Machine)來完成，使用了4500張影像來做實驗，其查準率(Precision)為98.34%，查全率(Recall)則是98.16%。 字元辨識的部分是使用類神經網路(Artificial Neural Network)中的反向傳播(Back Propagation)來辨識，讓內部自動反饋更新權重，使實際輸出值越來越接近目標輸出值，這個部分使用了5000張影像實驗，其準確率達到了99.07%，而在使用機器學習時，皆利用了逐次半自動遞迴標籤法，大量減少手動標籤的次數。

The purpose of this thesis was to study the method of image recognition for number tags. The major research includes plate localization and character recognition two portions. Plate localization was applied by lots of methodology, e.g. edge detection, erosion and dilation of morphology, support vector machine of machine learning. The localization experiment was executed by using 4,500 images and it was found that the precision is 98.34%, the recall is 98.16%. Character recognition was applied by back propagation of artificial neural network. It leveraged internal feedback to automatically update priority weights, therefore the mechanism could reach the target by operating more or more trials. The recognition experiment was executed by using 5,000 images and it was found that the accuracy is 99.07%. The method of applying machine learning was conducted on semi-recursive labeling to significantly reduce the cost of manual labeling.