201024716 六、發明說明： 【發明所屬之技術領域】 本發明係關於檢查晶圓圖案之圖案檢查方法及装置。 【先前技術】 自先前以來，已知例如如日本特開平5_28丨丨5丨所示之晶 圓之圖案檢查裝置，其具備：輸入晶圓之檢查圖像之機 構；比較所輸入之檢查圖像與預先記憶之原版圖像之機 構；由比較圖像之差異量判斷晶圓良否之機構；當前述差 ❿ #量在特定值以下時，將前述輸入之檢查圖像作為新的原 版圖像予以記憶之機構；及輸出前述良否之判斷結果之機 構。 。 但，先前之圖案匹配檢查中若圖案差異較大，則無法檢 • 查。其原因為，差異量過大，有判斷結果經常變為「否」 之虞。 另一方面，現狀也有因製造程序之不同而以圖案不同 • 之晶圓，即圖案偏差、顏色不均等不同之晶圓成為檢查對 象之情況。 〜另’也有因製品不同’而有圖案之差異不成為品質、性 月匕上問題之情形。如此之製品在晶圓圖案匹配檢查中無法 適當地檢查。 另方面，圖案不同之晶圓之情形下，關於差異，若使 利用學習處理之參考圖像之平均化與檢查靈敏度較寬鬆， 則可進行某種程度的檢查。 利用差異之學習處理之參考圖像之平均或檢測靈敏 143456.doc 201024716 度右過於寬鬆，則缺陷檢測性降低。 的缺陷檢測 並且’自動進行學習處理時，難以保證期望 性0 【發明内容】 本發明之課題為提供-種晶圓之圖案檢查方法及I置， 而可對於圖案不同之晶圓進行自動學習作用。 若例示用以解決前述問題之本發明之解決方式則為如 _下所示。 ⑴-種晶圓之圖案檢查方法，其特徵係具備下列步驟··輸 入作為檢查對象之晶圓之圖案或晶片之檢查圖 <象，比較所 輸入之檢查圖像與預先記憶之參考圖像，由比較圖像之差 異量，以判斷圖案或晶片良否； 其中，當檢查中良品率下降至一定的臨界值以下時利 用檢查中之圖案或晶片之圖像再度進行學習處理，作成新 的參考圖像，或以學習處理學習圖案後，尋找均句之圖 案，若為均勻之圖案部分，則以圖案匹配檢查以外之間距 檢查等之檢查進行晶圓之檢查，或以均勻之圖案部分同時 進行圖案匹配檢查與間距檢查，由間距檢查之結果決定圖 案匹配檢查之靈敏度，以決定之靈敏度檢查晶圓全面之圖 案或晶片。 (2)一種晶圓之圖案檢查裝置，其特徵係具備輸入作為檢查 對象之晶圓之圖案或晶片之檢查圖像之機構；比較所輸入 之檢查圖像與預先記憶之參考圖像之機構；具有基於比較 結果判斷晶圓良否之判斷機構之運算處理機構； 143456.doc 201024716 八中該運鼻處理機構，在檢查中良品率下降至一定的 匕二值以下時’利用檢查中之®案或晶>{之圖像再度進行 學％處理，作成新的參考圖像，或以學習處理學習圖案 後尋找均勻之圖案，若為均勻之圖案部分，則以圖案匹 配檢查以外之間距檢查等之檢查進行晶圓之檢查，或以均 勻圖案邛77同時進行圖案匹配檢查與間距檢查，由間距 檢查之結果決定圖案匹配檢查之靈敏度，以決定之靈丄 檢查晶圓全面之圖案或晶片。 即使於與原來之參考之圖案偏差、顏色斑驳有較大差異 ;!=距時，亦可判斷是否為作為晶圓不會成為製品 :广之良品，可保持檢查之作業效率、原來 之維持。 以下說明用以實施本發明之最佳形態。 ()檢查t大量地發生過度檢測時（良品率極端下降 ==學習處理。若詳述，則係以學習處理學習圖案後 句之圖案’若為均句之圖案部分， 之圖案匹配檢查以外之檢查，進行晶圓之檢杳。-等 ⑺ ：均勾圖案部分同時進行圖案匹配 查，由間距檢查之結果決㈣案匹配檢查之靈敏度。 片）以如前述決定之靈敏度檢查晶圓全面之圖案或晶 為不二T有較大差異之情形，亦可判斷是否 =:Γ= 能上問題之良品，而可保持檢杳 之作業效率、原來良品率之維持。 143456.doc 201024716 良品率極端地下降成為一定的臨界值以下時再次進行 學習處理’但其-定的臨界值’彳由裝置之使用者任意地 設定（例如90〜95%間之良品率）。其中，特定的臨界值可 僅固定於1個期望值’可僅固定最低值，亦可使期望值成 為可變而對應於必要加以調整。例如，亦可將最低值固定 為90%，設定其以上之任意臨界值（例如95%^尤其較好 可對應於檢查對象調整至期望之臨限值。 本發明係改良晶圓之圖案檢查方法及裝置者其具備下 列步驟.輸入晶圓之圖案或晶片之檢查圖像，比較所輸入 之檢查圖像與預；^5己憶之參考圖像，由比較圖像之差異 量，判斷晶圓良否。 本發明中，檢查中良品率極端地下降成為特定值（例如 90 /〇或95% ’或其他值）以下時，％用檢查中之圖案或晶片 之圖像再度進行學習處理，作成新的參考圖像，或以學習 處理學習圖案後，尋找均勻之圖案，若為均勻之圖案部 分，則以圖案匹配檢查以外之間距檢查等之檢查進行晶圓 之檢查，或以均勻之圖案部分同時進行圖案匹配檢查與間 距檢查，由間距檢查之結果決定圖案匹配檢查之靈敏度， 以決疋之靈敏度檢查晶圓全面之圖案或晶片 【實施方式】 實施例 參照附圖’說明本發明之較佳實施例。 圖1係顯示本發明中實施晶圓之圖案（外觀）檢查方法之 較適宜外觀檢查裝置1〇。 143456.doc 201024716 圖1 (A)中，外觀檢查裝置丨〇係用於判斷例如如圖1 (b)所 示之於半導體晶圓丨丨上排列形成之多數半導體晶片Ha中 所形成之各個電路圖案之缺陷是否在容許範圍内。以下， 以使用於檢查在半導體晶圓11上所形成之半導體晶片ua 為例說明本發明。 外嬈檢查裝置10，如圖丨（八）所示，具備光學攝影機構 1 〇a及控制該光學攝影機構之動作且用以對由該光學攝影 機構l〇a所得之圖像資訊進行運算處理之控制運算機構 10b ° 光學攝影機構10a’具有：設有保持半導體晶圓π之台 12a之移動部12 ;用以使該移動部之台123在χγ平面上， 於X軸方向、γ軸方向移動，繞著2軸旋轉之驅動器13 ;藉 由照明部14之照明下，用以對台12a上之於半導體晶圓η 上所形成之期望的半導體晶片lla之表面圖像進行攝影之 攝像部15。該攝像部15 ^如先前所知，以如CCD攝像元件 及其光學系統所構成。 控制運算機構（運算處理部）10b具有運算處理電路16， 忒運算處理電路16可藉由例如按照記憶於記憶體17之程式 動作之中央處理裝置（CPU)所構成。運算處理電路16，通 過控制電路18，控制光學攝影機構10a之驅動器13、照明 部14及攝像部15之各動作，另，依據記憶體17中所儲存之 貝訊對由攝像部15所得之圖像實施缺陷檢測處理。 "亥運算處理電路16’設有區域設定部16a，用以將藉由 攝像部15所得之圖像之檢查區域區劃成複數區域；缺陷抽 143456.doc 201024716 出。PI 6b，其藉由比較前述圖像之檢查區域與檢查用之模 版而抽出缺陷部位；判斷部16e，其判斷由該缺陷抽出部 所抽出之缺陷部位是否在容許内；學習機能部i6d，其在 為容許内但有作為製品無性能上問題之圖案偏差時，以其 偏差之圖案之圖像或晶片之圖像或以該等之平均圖像作為 參考圖像（參考圖像或基準圖像）加以學習。 另，運算處理電路16，連接有具有例如以液晶或cRT構 成之顯示部之終端機19，與以例如鍵盤及滑鼠等所構成之 輸入部20。終端機19可顯示以攝像部15攝影之圖像及以運 算處理電路16處理之圖像，且顯示必須操作光學攝影機構 1〇a之資訊。基於該等終端機I9中顯示之資訊，可自輸入 部20適當輪入必須操作外觀檢查裝置1〇之命令。 攝像部15對參考圖像（參考圖像或基準圖像）及被檢查體 用之圖像進行攝影。自由攝像部15所攝影之半導體晶片 1U之表面圖像，切出期望之檢查區域，顯示於終端機19 上。 圖2中，左侧之圖像係如前述切出之檢查區域之顯示晝 面21A之-例。該顯示畫面21A係顯示半導體晶片山係記 憶體晶片之例。 圖6顯示圖案圖像、晶片圖像。表面圖像中’於星號即 ★記號所示處觀察到缺陷部位。所觀察之缺陷為電路圖 案上之異物附著或電路圖案之部分缺損等之缺陷。 參考圖像（基準圖像），係自由半導體晶圓u之各半導體 晶片Ha所得之前述表面圖像21A中，選擇由缺損或異物所 143456.doc 201024716 成之前述缺陷較少之最優質半導體晶片lla之表面圖像 (21A)或平均圖像’以此作為參考圖像（基準圖像）而使用。 L參考圖像（基準圖像），與檢查對象即被檢查體之其他半 導體晶片Ua之表面圖像21Λ，藉由運算處理電路16加以比 較以抽出缺陷。 運算處理電路16，如先前所知，進行參考圖像（基準圖 象）/、被檢查體之半導體晶片1 i a之圖像比較。就圖案圖像 • 全體或圖案圖像中之一部分晶片圖像，就參考圖像（基準 圖像）與被檢查體之圖案全體或晶片，進行靈敏度調整。 此時，作為前處理，進行用以除去藉由照明部14引起之 ’’、、月不均之遮蔽處理；用以促進邊緣明確化之多值化處 理’用於減低邊緣檢測時圖案之粗密或圖像濃淡之影響之 色調變換處理；使用用以使識別圖案變容易之色度變換或 用以除去雜訊之膨脹/收縮過濾器之膨脹/收縮處理等。該 等處理’可以先前已知之方法實施。此等可適宜地選擇並 組合。 運算處理電路16 ’係對參考圖像（基準圖像）與被檢查體 之施加圖像前處理之圖像進行比較’由圖案之粗密或圖像 之濃淡、色度等，利用參考圖像（基準圖像）調整檢測靈敏 度。 以下’說明最合適之檢查順序之一例。 如圖2所示’於圖案匹配檢查中，事前登記成為比較對 象之圖像。 首先’作成成為檢查基準之參考圖像（基準圖像）(作成處 143456.doc 201024716 理方法）。士昧 a* . — 此Bf 參考圖像（基準圖像），如圖2之左側所 丁為例如1晶片。但亦可以是複數之晶片圖像為一組作 為基準圖像。 j於裝置之與運算處理部另外設置之記憶部（記憶體）中， 己L該參考圖像（基準圖像卜通過運算處理部之學習機能 進订學習。學習後，如圖2之右側所示，以平均圖像作為 /考圖像(基準圖像)並記憶於記憶部。 …、·σ果，如圖3所示，左侧之圖案由於與參考圖像（基準 圖像）一致，因此判斷為檢查ΟΚ，但右側之圖案，由於圖 案與參考圖像（基準圖像）不同，因此判斷為檢查NGe 由於圖案偏差，因此檢查之晶圓之圖案(或晶 发G，但稱微圖案偏差有於作為製品時不成為問題 所者。因此’若良品率（與基準圖像差異量較少之 良口口所佔之比例）下降 圖傻r其嘴 降至特疋值以下，則自動地重作參考 圖像（基準圖像）。例如， ^ 作# + A 〇良品率持續圖案檢杳 作業中，於良品率突 - 95% . S,/ ή L /〇Bf，若臨界值設定為 J自動地重作參考圖像（基準圖 開圖案檢查。 ）乂其良。口率再 如圖4所示，使左側之變更前之 學習，平均圖像變更& 成為右側之參考（再 J q诼變更）。例如良品率 為製品不成為門顳> 古& $ 降至95/。時，可將作 準圖像重作參考圖像。變更平 案（或4)作為基 (基準圖γ金、 象’再學習參考圖像 (基丰圖像），以此參考圖像（基準圖 :圖像 晶圓之圖案或晶片之檢查。 作為基本再次進行 i43456.cioc 201024716 如圖5所示，若修正參考圖像（基準圖像），則即使與參 考圖像修正前之圖案不同，檢查結果亦成為OK。 根據本發明，較好在例如良品率比最低之90%小，例如 為85%，或下降至其他特定值時馬上進行再學習處理重 作參考圖像（基準圖像）。 再者，本發明中，不限於一定的固定良品率（95%或其他 又疋值）’作業者可在特定的最低良品率（例如，但不 限於此）為止之範圍内以任意臨界值自由地再學習之方 式’調整學習機能。 又，較好具有以參考圖像（基準圖像）作為基本進行圖案 匹配檢查同時自動調整檢測靈敏度之機能之運算控制部。 此處所明靈敏度調整，意指比較圖案匹配檢查與間距檢 查之檢測性而自動決定靈敏度之機能。 如圖6所示，圖案匹配檢查，通常，如左側所示，使晶 片全體與參考圖像（基準圖像）之全體進行比較，進行一致 與否之檢查，但較好如右側所示，亦同時在檢查圖像内僅 對圖案中均勻之圖案部分進行比較檢查之間距檢查。 較好，比較圖案匹配檢查之結果與間距檢查之結果，以 可檢測相同缺陷之方式調整圖案匹配檢查之靈敏度。以相 同檢查圖像，進行圖案匹配檢查與間距檢查，以缺陷檢測 性能大致成為相同之方式，調整圖案檢查之檢測靈敏度。 再者，圖案匹配檢查雖可檢查晶片全體，但間距檢查通 常只檢查均勻圖案部分。 圖7顯示流程。 143456.doc 11 201024716 該圖示例中，以第一片晶圓作成檢查方法，進行晶圓檢 查。良品率比設定之值（例如90%)更低時，以檢查中之晶 圓進行學習處理。以再學習後之參考圖像（基準圖—像）為: 本進行再讀查。將再學㈣之參考圖像（基準圖像）之檢 查結果作為晶圓檢查結果加以保存’並結束檢查。 對於圖案產生偏差之晶片較好進行如下對應。 即使每個批次或每個晶圓圖案有偏差之情況，亦有可能 成為良品°圖案產生偏差後，對所發生之晶片重新學習處 理並進行重新檢查。 再者’以裝置自動判斷良品率下降之原因是否為圖案偏 差並不容易，但較佳為於良品率變成特定的臨界值以下後 重作學習處理並檢查。 本發明較佳之樣態為’良品率低於特定臨界值時重作學 習處理。 實驗例中，若良品率下㈣進行再次學習處理並檢查， 則可大幅抑制圖案偏差之過度檢出地加以檢查。 再者’依據晶圓而異，而有藉由學習之檢測性能下降 者。例如，顏色不均過甚之a m斗、# ± _ 仓您日日圓或於相同位置缺陷連續之 晶圓等’顯著地出現檢測性能之下降，但即使該等晶圓 中，與間距檢查併用，藉由進行檢測靈敏度之調整，可將 作為製品其性能上不成問題之圖案或晶片判斷為良品，而 可保持檢查之作業效率，維持原本之良品率。 【圖式簡單說明】 圖1(A)、（Β)係顯示本發明之晶圓之圖案檢查裝置之概 143456.doc -12- 201024716 略圖； )之圖； 之發生圖案 偏差 圖2係用以說明作成處理程式（參考 圖3係用以說明本發明中作為對象 片之圖； 圖4係用以說明藉由再學習作成參考圖像之圖； 圖5係用以說明再學習後檢查之圖； 圖6係用以說明圖案檢查與間距檢查差異之圖；及 圖7係顯示流程圖。201024716 VI. Description of the Invention: [Technical Field of the Invention] The present invention relates to a method and apparatus for inspecting a pattern of a wafer. [Prior Art] Since the prior art, a pattern inspection apparatus of a wafer, for example, as shown in Japanese Patent Laid-Open No. Hei No. 5-28丨丨5, is known, which has a mechanism for inputting an inspection image of a wafer, and compares the input inspection image. a mechanism for pre-memorizing the original image; a mechanism for judging whether the wafer is good or not by comparing the difference between the images; and when the amount of the difference is less than a specific value, the input inspection image is given as a new original image The organization of the memory; and the institution that outputs the judgment result of the aforementioned good or bad. . However, if the pattern difference is large in the previous pattern matching check, it cannot be checked. The reason is that the amount of difference is too large, and the judgment result often becomes "no". On the other hand, there are cases where wafers having different patterns due to different manufacturing processes, that is, wafers having different pattern deviations and uneven colors, are inspected. There is also a difference in the pattern between the other products because of the difference in the product. Such articles cannot be properly inspected in wafer pattern matching inspection. On the other hand, in the case of a wafer having a different pattern, if the averaging and inspection sensitivity of the reference image using the learning process are loose, the degree of inspection can be performed. The average or detection sensitivity of the reference image processed using the difference 143456.doc 201024716 The degree is too loose, and the defect detection is reduced. When the defect is detected and the learning process is automatically performed, it is difficult to ensure the expectation. [Invention] The object of the present invention is to provide a pattern inspection method and an I-type wafer, and to automatically learn the wafers having different patterns. . The solution to the present invention exemplifying the above problems is as shown below. (1) A method for inspecting a pattern of a wafer, comprising the steps of: inputting a pattern of a wafer to be inspected or an inspection image of a wafer; comparing the input inspection image with a pre-memorized reference image By comparing the difference amount of the image to judge whether the pattern or the wafer is good or not; wherein, when the good product rate in the inspection falls below a certain critical value, the learning process is performed again by using the image of the inspection or the image of the wafer to create a new reference. Image, or after learning to process the learning pattern, look for the pattern of the uniform sentence. If it is a uniform pattern part, check the wafer with the inspection of the interval check, etc., or simultaneously in a uniform pattern. The pattern matching check and the gap check determine the sensitivity of the pattern matching check by the result of the gap check, and determine the overall pattern or wafer of the wafer by determining the sensitivity. (2) A wafer pattern inspection device comprising: a mechanism for inputting a pattern of a wafer to be inspected or an inspection image of a wafer; and a mechanism for comparing the input inspection image with a previously stored reference image; An arithmetic processing mechanism having a judging mechanism for judging whether or not the wafer is good or not based on the comparison result; 143456.doc 201024716 The mid-nose processing mechanism of the eighth middle school, when the yield rate of the inspection is lowered to a certain value below the threshold value, 'using the inspection case or The image of the crystal >{ is again processed to a new reference image, or a learning pattern is learned to find a uniform pattern, and if it is a uniform pattern portion, the pattern matching check is performed in addition to the interval check, etc. Inspect the wafer for inspection, or perform pattern matching inspection and spacing inspection in a uniform pattern 邛77. The sensitivity of the pattern matching inspection is determined by the result of the spacing inspection to determine the overall pattern or wafer of the wafer. Even if there is a large difference between the pattern deviation and the color mottle of the original reference; if the distance is ~, it can be judged whether or not the wafer will not become a product: a good product can maintain the efficiency of the inspection and maintain the original. The best mode for carrying out the invention will now be described. () Check that t is excessively detected in a large amount (the yield is extremely low == learning processing. If it is detailed, the pattern of the sentence after the learning process is processed] is the pattern part of the sentence, and the pattern matching check is Inspection, wafer inspection. -etc. (7): The pattern is checked at the same time in the pattern of the hook pattern, and the sensitivity of the matching check is determined by the result of the gap check. The film is checked for the full pattern of the wafer with the sensitivity determined as described above. Or the case where the crystal is not the same as the T, it is also possible to judge whether or not =:Γ= can be a good product of the problem, and can maintain the efficiency of the inspection and the maintenance of the original yield. 143456.doc 201024716 When the yield rate is extremely lower than a certain critical value, the learning process is performed again, but the threshold value is set arbitrarily by the user of the device (for example, a yield of 90 to 95%). Here, the specific threshold value may be fixed only to one expected value 'only the minimum value may be fixed, or the expected value may be made variable to be adjusted as necessary. For example, the minimum value may be fixed to 90%, and any threshold value above it may be set (for example, 95%^ is particularly preferably corresponding to the inspection object adjusted to the desired threshold value. The present invention is an improved wafer pattern inspection method. And the device has the following steps: inputting the pattern of the wafer or the inspection image of the wafer, comparing the input inspection image with the pre-reviewed image, and judging the wafer by comparing the difference of the image In the present invention, when the inspection yield rate is extremely lowered to a specific value (for example, 90 /〇 or 95% ' or other value), % is again processed by the pattern or wafer image in the inspection to create a new one. The reference image, or after learning to process the learning pattern, to find a uniform pattern, if it is a uniform pattern portion, the wafer inspection is performed by inspection of the interval check or the like, or in a uniform pattern portion. Perform pattern matching inspection and spacing inspection, determine the sensitivity of the pattern matching inspection by the result of the spacing inspection, and check the overall pattern or wafer of the wafer with the sensitivity of the determination. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The preferred embodiment of the present invention will be described with reference to the accompanying drawings. Fig. 1 is a view showing a suitable visual inspection apparatus for performing a pattern (appearance) inspection method of a wafer in the present invention. 143456.doc 201024716 In A), the visual inspection device is used to determine whether, for example, the defects of the respective circuit patterns formed in the plurality of semiconductor wafers Ha arranged on the semiconductor wafer as shown in FIG. 1(b) are within the allowable range. Hereinafter, the present invention will be described by taking an example for inspecting a semiconductor wafer ua formed on a semiconductor wafer 11. The external inspection device 10, as shown in Fig. 8 (8), is provided with an optical imaging mechanism 1a and control The operation mechanism 10b of the operation of the optical imaging mechanism for calculating the image information obtained by the optical imaging unit 10a has an optical imaging mechanism 10a' having a stage 12a for holding the semiconductor wafer π. a moving portion 12; a driver 13 for moving the table 123 of the moving portion in the X-axis direction and the γ-axis direction on the χγ plane, and rotating around the two axes; and being illuminated by the illumination unit 14 The image pickup unit 15 for photographing the surface image of the desired semiconductor wafer 11a formed on the semiconductor wafer η on the stage 12a. The image pickup unit 15 is as previously known, and is constituted by, for example, a CCD image pickup element and an optical system thereof. The control arithmetic unit (arithmetic processing unit) 10b includes an arithmetic processing circuit 16, and the arithmetic processing circuit 16 can be constituted by, for example, a central processing unit (CPU) that operates in a program stored in the memory 17. The arithmetic processing circuit 16 passes The control circuit 18 controls the operations of the driver 13, the illumination unit 14, and the imaging unit 15 of the optical imaging unit 10a, and performs defect detection processing on the image obtained by the imaging unit 15 based on the information stored in the memory 17. The "Hai processing circuit 16' is provided with a region setting portion 16a for dividing the inspection region of the image obtained by the imaging portion 15 into a plurality of regions; the defect is 143456.doc 201024716. PI 6b, which extracts the defective portion by comparing the inspection region of the image with the template for inspection; the determination portion 16e determines whether the defective portion extracted by the defect extraction portion is within the allowable state; the learning function portion i6d, In the case of a pattern deviation which is not allowed to be a problem in the product, the image of the deviation pattern or the image of the wafer or the average image is used as the reference image (reference image or reference image) ) to learn. Further, the arithmetic processing circuit 16 is connected to a terminal unit 19 having a display unit composed of, for example, a liquid crystal or a cRT, and an input unit 20 composed of, for example, a keyboard and a mouse. The terminal unit 19 can display an image photographed by the image pickup unit 15 and an image processed by the operation processing circuit 16, and display information necessary to operate the optical photographing unit 1a. Based on the information displayed in the terminal unit I9, the command to operate the visual inspection device 1 can be appropriately rotated from the input unit 20. The imaging unit 15 photographs a reference image (reference image or reference image) and an image of the object to be inspected. The surface image of the semiconductor wafer 1U imaged by the free image capturing unit 15 is cut out of a desired inspection area and displayed on the terminal unit 19. In Fig. 2, the image on the left side is an example of the display surface 21A of the inspection area cut out as described above. This display screen 21A is an example of a semiconductor wafer mountain memory chip. Fig. 6 shows a pattern image and a wafer image. In the surface image, the defect is observed at the position indicated by the asterisk, the mark. The defects observed are defects such as adhesion of foreign matter on the circuit pattern or partial defects of the circuit pattern. In the aforementioned surface image 21A obtained by the semiconductor wafers Ha of the free semiconductor wafers, the reference image (reference image) is selected from the defect or foreign matter 143456.doc 201024716. The surface image (21A) or the average image ' of lla is used as a reference image (reference image). The L reference image (reference image) is compared with the surface image 21 of the other semiconductor wafer Ua of the object to be inspected, which is the object to be inspected, and is compared by the arithmetic processing circuit 16 to extract the defect. The arithmetic processing circuit 16 performs image comparison of the reference image (reference image) / the semiconductor wafer 1 i a of the object to be inspected as previously known. In the pattern image • The wafer image of one of the whole or the pattern image is subjected to sensitivity adjustment with reference to the image (reference image) and the entire pattern or wafer of the object to be inspected. At this time, as a pre-processing, a masking process for removing the '', and the moon unevenness caused by the illumination unit 14 is performed; a multi-valued process for promoting edge clarification' is used to reduce the coarseness of the pattern at the time of edge detection. The tone conversion processing of the influence of the image shading; the chromaticity conversion for making the recognition pattern easy, the expansion/contraction processing of the expansion/contraction filter for removing noise, and the like. These processes can be carried out in a previously known method. These can be suitably selected and combined. The arithmetic processing circuit 16' compares the reference image (reference image) with the image of the image to be processed before the image is applied, 'by the coarseness of the pattern, the shade of the image, the chromaticity, etc., using the reference image ( The reference image) adjusts the detection sensitivity. The following 'an example of an inspection sequence that is most suitable is described. As shown in Fig. 2, in the pattern matching check, the image is registered as a comparison object in advance. First, a reference image (reference image) to be the inspection standard is created (the creation method is 143456.doc 201024716). Gentry a* . — This Bf reference image (reference image), as shown on the left side of Figure 2, is for example a wafer. However, it is also possible that a plurality of wafer images are used as a reference image. In the memory unit (memory) provided separately from the device and the arithmetic processing unit, the reference image (the reference image is learned by the learning function of the arithmetic processing unit. After learning, as shown in the right side of FIG. 2 It is shown that the average image is used as the test image (reference image) and is memorized in the memory portion. ..., σ fruit, as shown in FIG. 3, the pattern on the left side is identical to the reference image (reference image). Therefore, it is judged that the flaw is checked, but the pattern on the right side is different from the reference image (reference image). Therefore, it is judged that the pattern of the wafer is inspected because of the pattern deviation (or the crystal pattern G, but the micro pattern). The deviation is not a problem when it is used as a product. Therefore, if the yield rate (the ratio of the good mouth with less difference from the reference image) decreases, the figure falls below the special value, and then the automatic Reproduce the reference image (reference image). For example, ^^# A 〇 yield rate continuous pattern check operation, the yield rate is -95%. S, / ή L /〇Bf, if the threshold is set Automatically redo the reference image for J (reference pattern open pattern check乂其良. The mouth rate is as shown in Fig. 4, so that the learning before the change on the left side, the average image change & becomes the reference on the right side (again J q诼 change). For example, the yield rate does not become a threshold. When the ancient & $ is reduced to 95/., the collated image can be re-referenced as a reference image. Change the case (or 4) as the base (reference map γ gold, image 're-learning reference image (Jifeng image) ), this reference image (reference map: image wafer pattern or wafer inspection. As a basic again i43456.cioc 201024716 as shown in Figure 5, if the reference image (reference image) is corrected, even if The result of the inspection before the correction of the reference image is different, and the result of the inspection is also OK. According to the present invention, it is preferable to perform the re-learning treatment immediately, for example, when the yield ratio is less than 90% of the lowest, for example, 85%, or when it is lowered to other specific values. As a reference image (reference image). Further, in the present invention, it is not limited to a certain fixed yield (95% or other depreciation) 'the operator may be at a specific minimum yield (for example, but not limited to Freely at any threshold within the range The learning method is to adjust the learning function. Further, it is preferable to have a reference image (reference image) as a calculation control unit that basically performs pattern matching inspection and automatically adjusts the detection sensitivity. The sensitivity adjustment here means the comparison pattern. The matching check and the detection of the pitch check automatically determine the sensitivity. As shown in Fig. 6, the pattern matching check is generally performed by comparing the entire wafer with the reference image (reference image) as shown on the left side. Check whether it is consistent or not, but it is better to check the interval between the inspections in the inspection image only in the inspection image. It is better to compare the result of the pattern matching inspection with the result of the spacing inspection. The sensitivity of the pattern matching check is adjusted in such a manner that the same defect can be detected. The pattern inspection check and the pitch check are performed in the same inspection image, and the detection sensitivity of the pattern inspection is adjusted in such a manner that the defect detection performance is substantially the same. Further, although the pattern matching check can inspect the entire wafer, the pitch inspection usually only checks the uniform pattern portion. Figure 7 shows the flow. 143456.doc 11 201024716 In the example of this figure, a wafer inspection is performed using the first wafer inspection method. When the yield is lower than the set value (for example, 90%), the learning process is performed by the crystal in the inspection. The reference image (reference map-image) after re-learning is: This is for re-reading. The inspection result of the reference image (reference image) of the re-learning (four) is saved as a result of the wafer inspection, and the inspection is ended. It is preferable to perform the following correspondence on the wafer in which the pattern is deviated. Even if there is a deviation in each batch or each wafer pattern, it is possible to re-learn and re-examine the generated wafer after the deviation of the pattern is good. Further, it is not easy to automatically judge whether the cause of the decrease in the yield rate is a pattern deviation, but it is preferable to repeat the learning process and check after the yield rate becomes equal to or lower than the specific threshold value. A preferred aspect of the invention is that the learning process is repeated when the yield is below a certain threshold. In the experimental example, if the re-learning process is performed and checked under the yield rate (4), the excessive detection of the pattern deviation can be greatly suppressed and inspected. Furthermore, 'depending on the wafer, there is a decrease in detection performance by learning. For example, if the color is not uniform, the #± _ warehouse, your yen, or the wafer with the same position defect, etc., will notice a significant drop in detection performance, but even if the wafers are used together with the spacing check, By adjusting the detection sensitivity, the pattern or wafer which is not a problem in the performance of the product can be judged as a good product, and the work efficiency of the inspection can be maintained, and the original yield rate can be maintained. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1(A) and FIG. 1 are diagrams showing a pattern inspection apparatus for a wafer of the present invention, 143456.doc -12-201024716; FIG. A description will be given of a processing program (refer to FIG. 3 for explaining a map as a target sheet in the present invention; FIG. 4 is a diagram for explaining a reference image by re-learning; FIG. 5 is a diagram for explaining a post-learning check. Figure 6 is a diagram for explaining the difference between pattern inspection and spacing inspection; and Figure 7 is a flow chart.
【主要元件符號說明】 10 外觀檢查裝置 10a 光學攝影機構 10b 控制運算機構 11 半導體晶圓 11a 半導體晶片 12 移動部 12a 台 13 驅動器 14 照明部 15 攝像部 16 運算處理電路 16a 區域設定部 16b 缺陷抽出部 16c 判斷部 17 記憶體 143456.doc 201024716[Description of main component symbols] 10 Visual inspection device 10a Optical imaging mechanism 10b Control computing mechanism 11 Semiconductor wafer 11a Semiconductor wafer 12 Moving portion 12a Stage 13 Driver 14 Illuminating unit 15 Imaging unit 16 Calculation processing circuit 16a Area setting unit 16b Defect extraction unit 16c Judging part 17 Memory 143456.doc 201024716
18 19 20 21A 控制電路 終端機 輸入部 畫面 143456.doc -1418 19 20 21A Control circuit Terminal Input section Screen 143456.doc -14