三維場景之文字辨識為近年來一項主要的研究課題。在過去，三維場景之文字辨識研究主要基於靜態影像之分析。然而，隨著科技的進步，手機攝影或行車紀錄器等設備已日漸普及於人類之生活中，因此使得動態影像之三維文字辨識的需求大幅提升，例如應用於招牌、廣告、地圖等實景文字之辨識，或是交通監控系統中之車牌辨識等等。然而，動態影像中之三維場景文字辨識最大之挑戰為影像輸入品質之不足，鏡頭容易因晃動而造成影像模糊及產生雜訊，也會造成文字辨識的問題。為了解決以上的問題，本計畫將建立一套智慧型的可以去除模糊化之三維文字辨識系統，能夠辨識移動中物體上之文字。利用拍攝的影片，再根據場景分析與文字特性將影像去模糊化，接著擷取出文字區塊的清晰影像再做辨識處理。本計畫為三年期之研究計畫。在第一年的工作中，我們將研究靜止攝影機拍攝移動物件的情形，進行去模糊化及文字區域的切割跟辨識。第二年則是處理移動攝影機針對靜止物體文字進行拍攝而產生的模糊現象，再進行去模糊化及文字切割辨識。第三年則是對一般化的情形，也就是取像設備與目標文字皆在移動的狀況進行辨識。最後，我們可以將三年的研究成果整合成一套完整之動態取像之文字切割與辨識系統。 3D real scene optical character recognition attracts the attention of many researchers which becomes an important research topic recently. In the past, 3D real scene character recognition lies mainly on the analysis of static image or picture. However, mobile hand-held devices or vehicle recoders have widely spread into our daily-life. Hence, the demand for the recognition of 3D characters in videos undergoing moving environments drastically increases. It can be employed in the recognition of characters embedded in road signs, buildings, or license plates in traffic surveillance systems. However, the challenges encountered in 3D moving environments are much more severe, such as poor image quality, noises generated due to the vibration or blurring of capturing devices, etc. They will more or less hinder the success of character recognition. To conquer these problems, we plan to develop an intelligent 3D real scene character recognition system especially devised for characters embedded in moving environments. In our system, it first extracts characters from moving videos. Then, delimit character blocks which contain clear character images through scene analysis and character characteristics for later recognition purpose. This project is a three-years project. In the first year, we will tackle the characters embedded in moving objects captured by static mobile devices. In the second year, the characters to be recognized are embedded in static objects captured by moving devices (such as cars). In the third years, we plan to extract and recognize the characters embedded in moving objects captured by moving devices. Moreover, the works accomplished in the three years will also be integrated to complete an intact 3D real scene character recognition system. 研究期間：10008 ~ 10107
