混合类神经网路应用於影像车号辨识.PDFVIP

International Journal of Advanced Information Technologies (IJAIT) 混合類神經網路應用於影像車號辨識 張財榮 賴建庭 南台科技大學 資訊工程系 南台科技大學 資訊工程系 trchang@mail.stut.edu.tw m94g0213@webmail.stut.edu.tw 摘要 本文呈現完整的影像車牌辨識流程，針對各個部分常見的問題提出改善方法。將車 牌辨識流程分為四大部分，分別為前處理、車牌區域定位、字元分割以及字元辨識。其 中車牌區域定位是基於模糊理論概念來完成，選取模糊集合中亮度較高的區域並加以評 選。在字元切割部份，文中將提出了快速且有效的字元分割方法。為了提升辨識效能， 我們採用混合類神經網路架構來進行辨識工作。其中自組織映射網路扮演了分派器以及 閘道器的角色，負責將輸入字元分派到若干個獨立的子網路，如此可降低每個子網路所 需辨識的類別數，進而達到提高辨識率的效果，由實驗結果發現可以提升百分之五的車 牌辨識率。 關鍵詞：車牌辨識、混合類神經網路、模糊理論、自組織映射網路、委員會機器。 Abstract In this paper, the complete procedure of image-based license plate number recognition was presented, and we addressed the improvement for the common problems in each parts of license plate recognition system. The whole procedure was divided into four sections that were separately preprocessing, locating of vehicle plate area, character segmentation and character recognition. The locating of vehicle plate area was achieved by the method base on fuzzy theory. The higher brightness areas in the fuzzy image were taken as the candidates for evaluating the exact license plate location. In the section of character segmentation, the fast, effective clustering algorithm and the tilt corrective algorithm were used for getting the correct and modified character connective components. Furthermore, in order to make the performance of character recognition higher, the module of mixture of experts was adopted. The self-organization map acted the distributor and gating network in this module, assigned overall classes to several experts called experts. Thus, the amount of classes which a experts in charged will be reduced and then the recognition rate will be promoted. After