【正文】 . Based on the generated 3D model, one can match the obtained model with the model in database, and the subtraction method will be carried out.Based on the range image, researchers have developed methods to extract features directly, especially for industrial purposes. Some of the methods are based on image information such as [24], and some methods are based on the geometric information the range image can provide. In [26], the author introduced a new feature extraction method using neighboring operators to detect the local feature of the range image. Others develop the feature extraction into a multiple feature groups and multiple views, based on which a 3D model can be reconstructed.Further more, with the assistance of the given design model, the feature detection from the range image will be much more effective and efficient. The authors in proposed a method using segmentation technique to recognize objects in the dense range image with the help of CAD model. To take advantage of the CAD model, researchers have developed algorithms to recognize features from the solid models .One of the primary targets of our system is to efficiently remove the feature lines, and thus to detect feature lines and model them is an important issue. For the modeling purpose, studies have been done as a subject of parameterization, or curve fitting. Authors of presented an algorithm based on physical phenomenon to partition a group of irregular points into subsets and fit a parameterized curve to each subset. presents a method to interpolate the points and fit with conic splines.All these research contributions provide us a solid foundation of our theory of dent detection and feature removal algorithms, so that our dent detection system can be designed on a reliable basis. Overview of our System DesignThe target of our automatic dent detection system is specified clearly in previous chapter, which is to find out the dent points in a range image, without the interference of noise and the existing feature lines of the model. The surface information is obtained by the 3D scanner, and discretized into a dense point clouds with unified x and y values – the range image. All our methods are based on this kind of data format.As stated in Chapter 1, an industrial car manufacturing system is a pipeline with processes from molding, part cutting, punching, plating, varnishing to assembling. The dents which are invisible until varnishing are most harmful to car part finishing, and most of the dents es from punching and other moving processes. Thus our automatic detection system should be integrated to the whole manufacturing line before plating and varnishing process. Also, there should be the least modifications to the existing system so that the performancecost ratio is maximized.Modern car manufacturing system is a very productive system producing thousands of cars each day. Since our target of the automatic dent detection system is to increase the detection accuracy and maximally reduce human interferences, so that the quality and efficiency will be highly increased, our integrated dent detection system should be very effective.Our target is to fully eliminate human involvement, therefore the detection quality should be very high, and the marked out dent areas should be exact, and no dents should be ignored for later on manual modification of the car parts.In order to meet these requirements, we introduce our automatic dent detection system with range image sensors, detection algorithm kernel embedded in PC and marking robot. The data is acquired by the range image sensor on a piecewise base, and the image sensor is sweeping across the whole car body area. The images are streamed to the detection kernel, and dent points with position information are output to the marking robot, the marking robot will then mark the dents according to the position information. The flow chart is as follows:附錄B中文翻譯汽車車身凹痕自動(dòng)檢測(cè)金屬板材的汽車車身有時(shí)會(huì)有小缺陷,這些缺陷非常難以被發(fā)現(xiàn)。在汽車制造業(yè),這些缺陷在鈑金汽車車身質(zhì)量是很有害的,但是,如果他們能夠在初級(jí)階段被自動(dòng)探測(cè)與去除，那么大量的修理工作的就可以省去。然而人類的眼睛很難發(fā)現(xiàn)小凹痕等缺陷, ,而且它的效率很低并且不準(zhǔn)確,因此利用其他手段的檢測(cè)是十分必要的,自動(dòng)化檢測(cè)系統(tǒng)能有效地檢測(cè)金屬板材表面缺陷和他們的位置,以便以后容易修理。1．測(cè)量的準(zhǔn)確度是一個(gè)很重要的指標(biāo),因?yàn)樵诎枷莸拇笮⊥ǔＪ欠浅Ｐ〉摹?．檢測(cè)速度必須是高效的。它被汽車制造商為提高生產(chǎn)的精度所廣泛接受,而不是一個(gè)長(zhǎng)時(shí)間等待檢測(cè)過(guò)程中,取出一些小型的缺陷。一個(gè)單一的金屬板材的汽車傳動(dòng)部位應(yīng)不超過(guò)10分鐘。3．這個(gè)系統(tǒng)應(yīng)該足夠智能,能確定一個(gè)凹痕對(duì)應(yīng)汽車的哪一部分。4．這個(gè)系統(tǒng)應(yīng)該需要的最少的人數(shù)參與。這種缺陷的檢測(cè)系統(tǒng)可以有兩種主要的結(jié)論。第一個(gè)是非常直截了當(dāng),采用表面信息數(shù)據(jù),。職位排除缺陷定位數(shù)據(jù)去哪里。另一個(gè)是從反向?qū)Ω哆@個(gè)問(wèn)題,它是檢查表面數(shù)據(jù)獲得,并找出所有的可能位置的凹痕,然后執(zhí)行一些分類和驗(yàn)證方法從而區(qū)分區(qū)分凹痕的功能,例如神器或噪聲和店員的設(shè)計(jì)模型。其馀的職務(wù)是缺陷的位置。第一種方法,雖然是非常直截了當(dāng),容易實(shí)現(xiàn),但還存在一些問(wèn)題,這將防止系統(tǒng)存在的魯莽性性和自動(dòng)性。因?yàn)檫@種方法基于減法,一個(gè)很好的曲面模型必須先推導(dǎo)。然而,由于表面檢測(cè)傳感器總是有大小限制,并沒(méi)有覆蓋整個(gè)模型,對(duì)齊、曲面重構(gòu)和表面局部登記是必要的,這是沉重的計(jì)算任務(wù)。另一個(gè)問(wèn)題是通常的生產(chǎn)線是一種振動(dòng)系統(tǒng),所以很難確定模型的精確定位,使登記是很困難的。因此一個(gè)獨(dú)立的模型系統(tǒng),最理想的是以最小的位置信息和登記的計(jì)算。第二種方法是基于局部特征提取,并在設(shè)計(jì)模型中扮演了一個(gè)重要角色。因?yàn)樗恍枰恢眯畔?沒(méi)有登記或重建是需要的。因此這種檢測(cè)方法越來(lái)越被廣泛應(yīng)用于各類缺陷的檢測(cè)和質(zhì)量檢驗(yàn)系統(tǒng)?，F(xiàn)代表面檢測(cè)系統(tǒng)采用各種各樣的方法,可有效無(wú)損傷表面檢測(cè)系統(tǒng),利用光學(xué)元件。常用的方法是物理光學(xué)方法。凹痕和波紋檢測(cè)問(wèn)題的最初是由二維成像為解決表面如烤漆表面閃閃發(fā)亮。目前世界汽車工業(yè)是發(fā)展和應(yīng)用技術(shù)的趨勢(shì)是不斷提高產(chǎn)品質(zhì)量、降低成本、重量和減少能源消耗和環(huán)境影響。在過(guò)去的十年中,機(jī)器視覺(jué)系統(tǒng)應(yīng)用緩慢，但成功應(yīng)對(duì)各種各樣的挑戰(zhàn)。自從我們的系統(tǒng)會(huì)被整合到管道前塞組件,反射的光將是一個(gè)嚴(yán)重的問(wèn)題。成像方法將會(huì)失敗,。在這種情況下,從上方看坑將完全看不見。利用物理光學(xué)方法的格局及密度的光反射出來(lái)的表面,是有嚴(yán)格的要求,分析了反射的光。在管道復(fù)雜的表面,有光線反射涂料,使該方法執(zhí)行起來(lái)很難。相比之下,機(jī)器視覺(jué)系統(tǒng)、三維激光掃描系統(tǒng)都有自己的優(yōu)點(diǎn):要求的也不是那么嚴(yán)格。只要有反射的光從表面反射出來(lái),掃描器能夠用傳感器反映所測(cè)距離。此外,不僅是表面的幾何信息,但還有其他的一些信息,如色彩、亮度,所有的2D圖像。因此我們使用這個(gè)方法可以為我們采集數(shù)據(jù),從表面,并利用一種特殊的形象——范圍影像幾何信息和圖像的信息。表面缺陷檢測(cè)的問(wèn)題已經(jīng)被研究了許多年。缺陷檢測(cè)技術(shù)經(jīng)常被使用在大多數(shù)工業(yè)領(lǐng)域,如薄片金屬成型,表面加工和拋光、快速成型、紡織設(shè)計(jì)、光盤設(shè)計(jì)、PCB及高品質(zhì)的芯片生產(chǎn)等。摘要當(dāng)前,在城市化發(fā)展優(yōu)質(zhì)表面制造在各個(gè)領(lǐng)域、表面缺陷在線檢測(cè)系統(tǒng),現(xiàn)在成為最重要的部分之一。當(dāng)前的缺陷檢測(cè)方法:表面質(zhì)量檢測(cè),傳統(tǒng)方法采用光學(xué)元件表面檢測(cè)。最古老的建筑物追溯到那些在邊緣檢測(cè)質(zhì)量的模式所形成的反射鏡反射光束從表面。運(yùn)用現(xiàn)代光學(xué)檢測(cè)系統(tǒng)的結(jié)構(gòu)光和激光掃描方法。反射的光路和密度是由表面狀況的檢查,然后對(duì)混凝土缺陷檢查,檢查的被反射的光模式。這些光學(xué)方法無(wú)疑是非常精確的,但是它同時(shí)也是顯而易見的,他們有很多的局限性。首先,他們擅長(zhǎng)平面表面上,但他們檢查檢驗(yàn)時(shí)通常會(huì)失敗的表面與特征。其次,很難獲得足夠的信息,因此從表面難以實(shí)現(xiàn)對(duì)缺陷診斷和可視化。第三,這些光學(xué)方法更多的依靠人類受牽累,他們其中的一些是每一個(gè)經(jīng)驗(yàn)和主觀的。很難實(shí)現(xiàn)自動(dòng)化缺陷檢測(cè)或整合到生產(chǎn)線。因此,這些光學(xué)方法應(yīng)用于區(qū)域需要更多和更少的效率更高的精度檢驗(yàn),如晶圓片。最近,最熱門的議題,自動(dòng)化表面缺陷的檢測(cè)是使用機(jī)器視覺(jué)。在該方法中,采用電荷耦合器件(CCD)的檢測(cè)傳感器。例如,SIMAC由Zumbach是一個(gè)表面質(zhì)量檢測(cè)系統(tǒng),利用CCD和圖像處理。以CCD攝像機(jī)、圖像檢測(cè)機(jī)器人檢檢查表面一塊一塊的,然后統(tǒng)計(jì)數(shù)據(jù)的處理與圖像處理的方法來(lái)檢測(cè)的缺陷。然而,這臺(tái)機(jī)器視覺(jué)方法有其不足之處。首先,精度是制約CCD像素?cái)?shù)字上。其次,由于缺乏幾何數(shù)據(jù)、圖像處理方法在某些特殊的情況下不能檢測(cè)缺陷。今天,越來(lái)越多的要求是提高缺陷的檢測(cè)。表面缺陷自動(dòng)檢測(cè)系統(tǒng),必須滿足下列要求。檢測(cè)系統(tǒng)的精度必須高,及三維測(cè)量數(shù)據(jù)必須被獲得。同時(shí),分析