【正文】 maximal stable extremal region (MSER). Using mathematical expression, that means if and only if ，gets the local minimum, MSER is obtained. The symbol denotes the step length of the gray level threshold in this formulation. Considering that gray levels can be adjusted in two opposite directions, after these two operations we can get two kinds of extremal regions respectively, ., bright extremal regions and dark extremal regions. The difference of these two regions also contributes to the inference of license plate location, as described in next subsection. The standard MSER detection is described above. In our implementation, linear time MSER detection [9] is utilized to increase the efficiency of MSER extraction. Compared with other region features, MSER has many advantages and can get better performance in most applications [10]. In our task of detecting characters, the main superiority is MSER’s invariance to lighting change. As long as the luminance in the image changes monotonously, the MSER can keep stable even illumination changes from daytime to nighttime. In order to detect blobs even in some lowcontrast image regions and ensure any MSER not tobe left, we slightly adjust the MSER detection process. First, we set the threshold step as a minimum value of 1. Then the diversity of extremal regions’ area in the same position is set to a very small value, so that we can get as many extremal regions as possible. Afterward, we only save the extremal regions with the minimum area variation among all regions in the same position as the final MSER. Meanwhile, the MSER are restricted by some preset license plate parameters. The remaining MSER can be considered as candidate license plate characters. An example of MSER extraction result in a test image is shown in Fig. 4. B. License Plate Location InferenceIf the extracted MSERS just are the characters in a license plate, we can infer the exact location of the license plate easily. However, in practical applications, just as the result shown in Fig. 4, some characters may not be detected as MSERS. The reason is various, such as the unequal illumination on license plate or the joint of characters with the license plate boundary. In addition, there are a large number of MSERS that meet the restriction in the background. Therefore, after we detect the MSERS in the image, we must check them and find those having similar layout with characters in standard license plates. The process of inference and analysis is described below.The concepts of nodes and edges in the graphical model are introduced in our inference. We intend to set the candidate characters as nodes and build an edge between two nodes which meet the geometric relationship and gray level relationship of two adjacent characters. First of all, we calculate the value of geometric relation and gray level relation between every two MSERS that were extracted before. Hence, the MSERS relationship matrix is built. The geometric relationship includes Euclidean distance, horizontal distance, vertical distance between two MSERS, difference of height and width between their bounding boxes, etc. The gray level relationship means that whether both blobs are bright or dark MSERS. Next we deal with each MSER to search in the left direction for another one that satisfies the geometric relation between two adjacent characters according to the value in the relationship matrix. If such is not found due to some undetected characters, then we search for two MSERS that have space of one character. Besides, the pair of matched MSERS must be the uniform bright or dark MSERS simultaneously. Then such two blobs can be set as nodes, and an edge is built between them. Meanwhile, the type of edge is labeled as either adjacent or at interval. The result of setting nodes and edges of Fig. 4 can be seen in Fig. 5, where yellow dots denote nodes and yellow。 license plate recognition。(2) 字符識(shí)別采用的是模板匹配算法，雖然字符模板都是標(biāo)準(zhǔn)的字符，然而在實(shí)際情況下由于各種原因?qū)е挛覀兦懈畹玫降淖址皇悄敲礃?biāo)準(zhǔn)，如字符缺損、傾斜、斷裂、粘璉等情況導(dǎo)致在匹配過(guò)程中會(huì)出現(xiàn)匹配錯(cuò)誤，尤其是在數(shù)字和英文字母的匹配識(shí)別過(guò)程中，例如數(shù)字1不能很好的識(shí)別、C和0、M和N，B和D、G和0等之間的混淆。最后當(dāng)所有模塊都按相同的算法與待識(shí)別圖像匹配完，具有最大匹配度的匹配則為最佳匹配，視為最終的匹配結(jié)果。(3)提出了一種基于改進(jìn)的彈性模板匹配法對(duì)車(chē)牌字符進(jìn)行識(shí)別。最后用投影法實(shí)現(xiàn)對(duì)車(chē)牌區(qū)域的精確定位。(2)提出了一種基于輪廓檢測(cè)和分級(jí)判斷的車(chē)牌定位算法。本文所設(shè)計(jì)的系統(tǒng)利用目前方興未艾的嵌入式硬件開(kāi)發(fā)技術(shù)和性能卓越的開(kāi)源嵌入式Linux操作系統(tǒng)及相關(guān)開(kāi)發(fā)工具，結(jié)合高效的車(chē)輛牌照自動(dòng)識(shí)別核心算法，系統(tǒng)穩(wěn)定快速可靠并具有很高的性能價(jià)格比，在ITS領(lǐng)域?qū)⒂兄鴱V泛的應(yīng)用前景。 正常情況識(shí)別結(jié)果 傾斜及陰暗情況識(shí)別結(jié)果一 傾斜及陰暗情況識(shí)別結(jié)果二 其他情況 其他識(shí)別一 其他識(shí)別結(jié)果二 其他識(shí)別結(jié)果三結(jié)論本文全面地評(píng)述了車(chē)牌定位與識(shí)別技術(shù)的研究現(xiàn)狀和發(fā)展概況。點(diǎn)擊切割按鈕后，最下面的7個(gè)小label會(huì)顯示出切割好并歸一化為20*40的車(chē)牌字符圖像 切割界面(5) 識(shí)別。點(diǎn)擊預(yù)處理按鈕，在label中會(huì)顯示經(jīng)過(guò)HSV分割，濾波，形態(tài)學(xué)處理后的圖像 預(yù)處理界面(3) 定位。打開(kāi)的照片會(huì)顯示在Label中。點(diǎn)擊讀取照片下拉框，里面有三個(gè)選項(xiàng)：固定路徑、自由選擇和攝像頭導(dǎo)入圖像。該算法能很好解決實(shí)際應(yīng)用中的車(chē)牌圖像字符變形、缺損等情況帶來(lái)的問(wèn)題。在車(chē)牌字符識(shí)別方面，本文提出一種基于改進(jìn)的彈性模板匹配法對(duì)車(chē)牌數(shù)字字母進(jìn)行識(shí)別。試驗(yàn)表明，該算法對(duì)光照不均勻、字符和底色對(duì)比度低以及車(chē)輛分布情況的復(fù)雜性等并不敏感，能夠快速準(zhǔn)確的定位出多車(chē)牌區(qū)域。 識(shí)別結(jié)果 本章小結(jié)本章研究了車(chē)牌定位、車(chē)牌字符分割和字符識(shí)別。彈性模板匹配算法是：先將每一個(gè)模板在待識(shí)別圖像中的某個(gè)小范圍內(nèi)(上下左右)進(jìn)行適當(dāng)?shù)囊苿?dòng)(約2—3像素)，來(lái)找到最大的匹配度。但是實(shí)際的車(chē)牌字符一般多少有一些變形和缺損，這就增加了正確識(shí)別字符的難度。與感知器和線性神經(jīng)網(wǎng)絡(luò)不同的是，BP網(wǎng)絡(luò)的神經(jīng)元采用的傳遞函數(shù)通常是Sigmoid型可微單調(diào)遞增函數(shù)．可以實(shí)現(xiàn)輸入到輸出的高度非線性映射。(2) BP神經(jīng)網(wǎng)絡(luò)識(shí)別算法。 識(shí)別算法概述(1) 模板匹配法。而車(chē)牌的字符包含了漢字、字母和數(shù)字,結(jié)構(gòu)比較復(fù)染,相似的字符較多,并且漢字筆畫(huà)多,非常復(fù)染,同時(shí)字符識(shí)別的算法也非常多,能夠清楚的識(shí)別每個(gè)字符是一個(gè)比較難解決的問(wèn)題[18]。 字符歸一化為了方便后續(xù)的字符識(shí)別操作，我們需要對(duì)分割好的字符進(jìn)行歸一化處理，我們把單個(gè)字符圖像歸一化為20*40大小的圖像。實(shí)驗(yàn)結(jié)果表明本文采用的算法能夠較好的實(shí)現(xiàn)單字符的分割，受字符粘結(jié)和斷裂影響不大。以前面的到得字符位置可以大概確定單個(gè)字符的寬度，再結(jié)合投影掃描法依次確定后續(xù)字符位置[17]。以第二個(gè)字符的開(kāi)始位置起點(diǎn)向左掃描，當(dāng)出現(xiàn)連續(xù)的兩列百點(diǎn)數(shù)大于5時(shí)，則認(rèn)為該列是第一個(gè)字符結(jié)束位置，繼續(xù)向前掃描，知道發(fā)現(xiàn)白點(diǎn)數(shù)小于2的那列，并把該列做為第一個(gè)字符的起始點(diǎn)。同理找到第3個(gè)字符的起始和結(jié)束位置。從第20列開(kāi)始向右掃描，直到遇到出現(xiàn)連續(xù)的8列白點(diǎn)數(shù)小于2時(shí)，再向右掃描遇到的白點(diǎn)數(shù)大于2的那列標(biāo)記為第二個(gè)字符的開(kāi)始位置。在分割前先估算出字符寬度、間距，并從圖像豎直方向的黑白跳變數(shù)定出字符的開(kāi)始位置，再參照估算值逐個(gè)定位切分全部字符。雖然大部分圖像都不可能達(dá)到這種質(zhì)量，但垂直投影仍然是字符分割中的一個(gè)非常重要的特征。 切割對(duì)預(yù)處理后的圖像采用投影法[16]，通過(guò)計(jì)算第x列上的白點(diǎn)數(shù)目，就可以得到x位置的垂直投影V(x)。 預(yù)處理分割車(chē)牌之前仍然需要對(duì)車(chē)牌進(jìn)行預(yù)處理，因?yàn)?，?chē)牌定位和矯正得到只是車(chē)牌的準(zhǔn)確位置，要提取字符，必須先確定字符的位置，但是字符的位置是相對(duì)車(chē)牌底色背景區(qū)域而言的，而且受到光照和車(chē)牌污點(diǎn)的干擾，所以必須進(jìn)行預(yù)處理。針對(duì)上述問(wèn)題,本文提出了基于簡(jiǎn)單統(tǒng)計(jì)法及投影法相結(jié)合的方法，對(duì)于具有復(fù)雜背景的字符圖像進(jìn)行二值化，并在此基礎(chǔ)上進(jìn)行形態(tài)開(kāi)閉運(yùn)算、進(jìn)行連通域判斷等方法去除二值圖的噪聲，同時(shí)保證字符的完整性，為最終進(jìn)行字符自動(dòng)識(shí)別提供了令人滿(mǎn)意的預(yù)處理效果。而聚類(lèi)法分割效果較好,解決了漢字不連通的問(wèn)題,但程序設(shè)計(jì)復(fù)雜、分割速度慢。其中行程標(biāo)記法的計(jì)算復(fù)雜度會(huì)隨著圖像幅度和連通域個(gè)數(shù)的增加而增長(zhǎng)得非?？?。本章先根據(jù)車(chē)牌的字符寬度、字符間距和排列規(guī)則等特點(diǎn)，對(duì)車(chē)牌字符進(jìn)行有效準(zhǔn)確的分割。 字符切割字符分割的任務(wù)是把多列或多行字符圖像中的每個(gè)字符從整個(gè)圖像中切割出來(lái)成為單個(gè)字符[14]