JP4220595B2 - Defect classification method and teaching data creation method - Google Patents

Description

【００３８】
なお、｜ＣｋＣｈ｜は、入力された教示点座標Ｃｈと既に入力された教示点座標Ｃｋとの間のユークリッド距離を表す。そしてＣｇは、入力された教示点座標Ｃｈに最も近傍の既入力教示点である。図１０に示す場合には、入力された教示点座標Ｃｈに最も近傍の既入力教示点ＣｇはＣ１となリ、ユークリッド距離がしきい値を満たすことになる。
【００３９】
【数５】

【００４０】
このように、ステップＳ６５において、ホストコンピュータ４０が、入力された教示点Ｃｈの近傍には既入力教示点Ｃｇが存在する（入力チェックがアクティブである＝ＹＥＳ）と判定した場合、ステップＳ６６において、ホストコンピュータ４０は、モニタ３８に対して図９に示すように探索した最近傍の欠陥番号ｇの教示用欠陥画像と入力された欠陥番号ｈの教示用欠陥画像とをカテゴリと共に表示する。なお、入力された欠陥番号ｈの教示用欠陥画像に対するカテゴリの付与は、ステップＳ６４において行われている。このように表示する教示用欠陥画像は、最近傍の教示用欠陥画像のみではなく、入力教示点Ｃｈとの距離の小さい順に操作者の指定する枚数だけ表示してもよい。次に、ステップＳ６６において、オペレータが、ステップＳ６４において付与されてカラム４７に表示された欠陥番号ｈの教示用欠陥画像のカテゴリと、既登録での最近傍の欠陥番号ｇの教示用欠陥画像のカテゴリとを比較し、その比較結果、カテゴリが同じ場合（別カテゴリの別教示点が近傍にない場合）は、ステップＳ６４において入力された教示点Ｃｈに対してカテゴリ一覧表を基に付与されたカテゴリがそのまま登録されることになる。もし、比較した結果、カテゴリが相違する場合、ステップＳ６７において欠陥番号ｈの教示用欠陥画像を表示領域４５に表示し、ステップＳ６８においてこの教示用欠陥画像を教示用として登録する必要がないと判断した場合、ステップＳ６２に戻って登録取りやめボタン６５を操作することによって登録を取りやめる。これによって、ステップＳ６９に進み、次の教示用欠陥画像が画像記録装置３４から呼び出されることになる。また、ステップＳ６８においてこの教示用欠陥画像を教示用として登録する必要があると判断した場合、カテゴリ表示カラム４７を用いて最近傍の欠陥番号ｇの教示用欠陥画像のカテゴリに合わせたカテゴリを入力することによって登録されることになる。
【００４１】
次に、ステップＳ７２において、教示用欠陥画像の全てに対してカテゴリが付与されるまで、繰り返されて終了となる。
即ち、予め設定した複数の特徴量若しくは複数の教示用欠陥画像セットから抽出される複数の特徴量を基準特徴量として用い、この基準特徴量により形成される多次元の特徴量ベクトル空間を用い、教示用欠陥画像を基準特徴量により特徴量ベクトル空間の１点に対応付け、同様にして複数の教示用欠陥画像を特徴量ベクトル空間に教示点として、それに対応する欠陥画像データが属する欠陥分類属性（カテゴリ）を対応付けして記憶する。このように、ホストコンピュータ４０が特徴量ベクトル空間に教示点Ｃｈをカテゴリを対応付けして記憶装置４３に教示用データとして登録する際、図９に示したモニタ３８の画面をとおし、オペレータは、ステップＳ６５およびＳ６６において表示領域４５に表示された入力された教示用欠陥画像を元に、表示領域５７およびカラム５８に表示された近傍教示用欠陥画像及びそのカテゴリを検討し、ステップＳ６８において入力された教示用欠陥画像の登録を既に入力したカデゴリで行なうか否かを決め、否の場合はステップＳ６２において登録を続行するか否かを決め、ステップＳ６８においてカテゴリを変えて登録をする場合は、ステップＳ６４と同様にカテゴリ一覧から選択あるいはステップＳ７０およびＳ７１と同様に新規カテゴリ追加の処理をもう一度実行し登録を行なう。
【００４２】
このように、ホストコンピュータ４０は、モニタ３８との間においてステップＳ６３〜Ｓ６８、Ｓ７０、Ｓ７１の処理を実行することによって、特徴量ベクトル空間においてカテゴリをきちんと区分けた形態で教示用データを取得することが可能となる。例えば、図５（ｂ）に示すように、２つの特徴量ベクトル空間において、×（カテゴリ１）、□（カテゴリ２）でもって、カテゴリをきちんと区分けた形態で教示用データを取得することが可能となる。
ところで、設定登録されたカテゴリは、欠陥番号など欠陥画像のインデックスとなる情報とともにホストコンピュータ４０内、あるいはホストコンピュータに接続された情報記憶手段（記憶装置）４３、あるいはネットワーク４２を介して接続されている前述の上位システムに教示用データとして記録される。教示用データのデータ内容は図８に示す通りである。特徴量の数ｍは、具体的に後述する特徴量の第１の実施例における予め設定した特徴量の数、あるいは特徴量の第２の実施例における固有空間の次元数ｓで決まる数である。
さらに、モニタ３８に表示される内容を図９に示す。４５は入力欠陥画像を表示する表示領域であり、４７は入力欠陥画像に対する入力カテゴリを表示するカラムである。５７は近傍既入力欠陥画像を表示する表示領域であり、５８は近傍既入力欠陥画像のカテゴリ、近傍既入力欠陥画像と入力欠陥画像のユークリッド距離、またその距離がしきい値（ＴＨｄ）以下の条件を満たすか否かのステータスフラグなどを表示するカラム、５９は距離順に表示された近傍既入力欠陥画像を閲覧する為、欠陥画像をスクロールするボタン、６０は再探索実行指示ボタン、６１は探索欠陥画像数入力カラム、６２はカテゴリ一覧表示ボタンあるいはカテゴリ一覧表示そのもの、６３は新規登録ボタンあるいは図７の４８、４９、５０、５１に示す新規カテゴリ入力に必要なボタン及び入力表示カラムを表す。４７に示す入力欠陥画像のカテゴリ表示カラムは入力も受け付け、カラムへの直接入力、あるいは６２、６３からの入力を介しての入力も受け付ける。６４は入力欠陥画像に対するカテゴリ登録を実行するボタン、６５は登録取りやめボタンである。
【００４３】
次に分類過程について説明する。ステップＳ３６において、画像処理装置３５は、画像記録装置３４に分類対象として入力されて記録された欠陥画像について同様にして、前述の予め設定したｍ個の特徴量の組み（ｃ1，ｃ2，…，ｃm）に基づく特徴量ベクトル（以降、分類データと呼ぶ。）として抽出し、該抽出された特徴量ベクトルをホストコンピュータ４０に提供し、記憶装置４３に記憶する。記憶される分類データは、図４（ｃ）に示すデータと同様な内容であるが、カテゴリ欄３６が空欄である。このため、分類用各欠陥について、その特徴量からカテゴリを前述の分類パラメータ（教示用データ）２７を用いて推定する。推定されたカテゴリは、ホストコンピュータ４０に接続された記憶装置４３に記憶される。記憶（記録）された分類データ３０は、ネットワーク４２を経由して上位システムである品質管理システム７に転送されることになる。
次に、ステップＳ３７において行う分類演算（分類データの作成処理）について図１１、図１２、および図１３を用いて具体的に説明する。まず、ホストコンピュータ４０は、分類対象とする入力された各欠陥画像の特徴量ベクトルを記憶装置４３から呼び出すことによって、分類対象の各欠陥画像の特徴量ベクトル空間を形成することができる。以下、分類対象の欠陥画像により生成された特徴量ベクトルにより表される特徴量ベクトル空間内の点を分類対象点Ｃと呼ぶことにする。
ステップＳ１０１において入力された分類対象である欠陥番号が入力欠陥画像についてホストコンピュータ４０は、ステップＳ１０２において分類対象点Ｃの最近傍にある教示点Ｃｅを探索し、該探索教示点Ｃｅに対応する教示用欠陥番号ｅに記録されている欠陥カテゴリを分類結果とすることで実行される。
【００４４】
学習過程において収集された教示用欠陥画像に対して算出された欠陥番号ｉの特徴量ベクトル空間におけるＣｉを教示点座標とし、分類過程において収集された分類対象の欠陥画像よりステップＳ３６において算出された特徴量ベクトルすなわち特徴量空間におけるＣを分類対象点座標とする。また、学習過程において収集された教示用欠陥画像に対してステップＳ３２において欠陥番号ｉに対応して付与されたカテゴリをＣａｔ（ｉ）、分類過程において収集された分類用欠陥画像に対してステップＳ３７で分類演算されるカテゴリをＣａｔと記すことにすれば、上述の分類処理は次に示す（数６）式にて記述することができる。
なお、教示用データとしては、ｎ個の教示用欠陥画像で構成されているものと仮定する。
【００４５】
【数６】

【００４６】
｜ＣｋＣ｜は教示点座標Ｃｋと分類対象点座標Ｃのユークリッド距離を表す。具体的には次に示す（数７）式で表される。
【００４７】
【数７】

【００４８】
（数６）式により、任意の分類対象はいかなるケースにおいてもそのカテゴリを一意に決定することができる。即ち、（数６）式の関係から、分類対象点座標Ｃを有する分類対象欠陥画像のカテゴリＣａｔは、教示点座標Ｃｋと分類対象点座標Ｃのユークリッド距離｜ＣｋＣ｜が最小を示す教示用欠陥番号ｉに付与されたカテゴリＣａｔ（ｉ）として一意に決定することができる。
【００４９】
しかしながら、分類対象点Ｃがいくら最近傍点であっても既登録の教示点Ｃｉから離れているものは既カテゴリには登録されていない、すなわち新たな欠陥カテゴリに分類されるべき未知欠陥である可能性もある。そこで、ステップＳ１０２において、ホストコンピュータ４０は、未知欠陥については既カテゴリに分類が不可能とし、操作者にカテゴリの確認を行なうのが妥当である。このような未知欠陥を、ホストコンピュータ４０において識別するために、教示点座標Ｃｉと分類対象点座標Ｃのユークリッド距離に対するしきい値ＴＨｄが入力手段４３等を用いてホストコンピュータ４０に対して設定される。次に、ホストコンピュータ４０は、この設定されたしきい値ＴＨｄと、（数３）式において算出される最小値ｍｉｎ｜ＣｋＣ｜とを、次に示す（数８）式に示すように比較してカテゴリは決定できるか否かを判断し、カテゴリが決定できる場合には、ステップＳ１０３においてカテゴリを記憶装置４３に格納される分類結果テーブルに登録し、カテゴリが決定できない場合には、ステップＳ１０４においてこの分類対象の欠陥番号ｉを記憶装置４３に格納される分類未知欠陥リストに追加する。
分類未知欠陥とは、記憶装置４３に記憶されている現在の教示用データから、分類対象の欠陥画像ｉが、特徴量ベクトル空間において、教示済の欠陥画像から大きく離れて存在してカテゴリを付与することが難しいものを示す。
【００５０】
【数８】

【００５１】
これにより、特徴量ベクトル空間において分類対象点Ｃからしきい値ＴＨｄ以内に（近傍に）教示点Ｃｉが存在しない場合には、分類対象点Ｃにおける分類対象の欠陥画像は属性不明、すなわち未知欠陥として分類することが可能となる。
【００５２】
次に、ホストコンピュータ４０は、ステップＳ１０５において、全ての分類用欠陥画像について分類カテゴリＢが付与されるまで、ステップＳ１０６に戻って次の分類用の欠陥画像が画像記録装置３４から読み出して繰り返すことになる。
更に、ホストコンピュータ４０は、ステップＳ１０７において、分類未知欠陥リストに欠陥番号が登録されているかを調べ、登録されていない場合は終了となり、登録されている場合には、ステップＳ１０８において未知欠陥に分類された入力欠陥画像についてオペレータのレビューにかけるために、未知欠陥に分類された入力欠陥画像（カテゴリ不明の欠陥画像）を図２に示すモニタ３８に表示させる。そして、ステップＳ１０９において、オペレータは、必要に応じて、未知欠陥のカテゴリを入力して付与する。カテゴリがオペレータにより提示された場合には、この未知欠陥に対し提示されたカテゴリを未知欠陥の特徴量ベクトルに付し、教示用データとして記録する。
この教示用データヘの記録処理は、ステップＳ１１０において、図４（ｃ）に示した教示用データリストに、新たな欠陥番号を追加し、追加された欠陥番号に対しカテゴリ、および特徴量ベクトル等を記録することにより終了となる。
【００５３】
即ち、ステップＳ１１０における分類未知欠陥へのカテゴリ付与について、図１２及び図１３に示すモニタ表示に沿って以下詳述する。まず、カテゴリ付与対象となる分類未知欠陥画像の選択は、図１２に示す一覧表示されている分類未知欠陥画像の中からユーザが選択指示することにより実行される。あるいは、キーボード４１による欠陥画像ファイル名入力、欠陥画像ファイル名リストからの選択でもよい。モニタ３８には、分類未知欠陥画像が表示領域６６に、特徴量ベクトル空間内での近傍点にある教示用欠陥画像がそのカテゴリと共に距離順に表示領域６７およびカラム７５に表示され、操作者は表示された近傍点教示用欠陥画像から類似する画像を選択し、カラム７３に分類未知欠陥のカテゴリを入力する。なお、ボタン７６は、距離順に表示された近傍点教示用欠陥画像をスクロールさせるためのものである。
【００５４】
ところで、近傍点教示用欠陥画像に類似する欠陥画像が無い場合は、表示領域７６に表示されたカテゴリ一覧表を参照し、該当カテゴリを選択するか、あるいは新規カテゴリを設定し、分類未知欠陥に新カテゴリを与え登録する。カテゴリ一覧表からの選択、あるいは新カテゴリ登録は学習過程で述べた方法と同一である。即ち、オペレータが、新規カテゴリ登録７７を指示することによって、キーボード４１の操作による入力カラム８０への新規カテゴリ情報の入力がホストコンピュータ４０によって実行される。入力の完了は７８を指示、入力の中断、取り消しは７９を指示する。また、欠陥画像詳細情報欄７４も学習過程で述べたもの５２と同一である。
最後に、分類登録のみのボタン６９により分類未知欠陥画像に対するカテゴリ入力を確定する。これに対し、登録ボタン６８は分類未知欠陥画像に対するカテゴリ入力を確定するとともに、この分類未知欠陥画像を教示用画像として、その画像と、それより得られる特徴量を新たに学習データ（教示用データ）に追加する。また、ステップＳ１１０における分類未知欠陥に対するカテゴリ分類放棄は削除ボタン７０で実行する。また、ボタン７１によって複数ある分類未知欠陥画像をリスト記録順に前後に呼び出す。すでにカテゴリが入力されていれば、そのカテゴリも表示される。ボタン７１の誤操作によって別の分類未知欠陥画像の呼び出しが行われた場合には、カテゴリ入力作業中の分類未知欠陥画像に対してはカテゴリは決定されずカテゴリ分類未実施のままとする。入力チェックボタン７２の機能については学習過程で述べたとおりである。
【００５５】
ところで、新たに追加される欠陥番号、対応するカテゴリ、特徴量ベクトルをまとめて新規教示用データと呼び、新規教示用データの特徴量ベクトルにより表される特徴量ベクトル空間内の点を新規教示点と呼ぶことにする。欠陥分類は、分類対象点Ｃが前述したように教示点Ｃｉの近傍に分類対象点がある場合、その分類対象の欠陥画像に対してその教示点のカテゴリを付与して行う。よって、新規教示点を追加したとしても、以前から存在する教示点の特徴量ベクトル空間内における位置は、新規教示点の追加によって変化しないので、新規教示点を追加以前の分類性能を保持することが可能となる。
例外は、既存の教示点の２×ＴＨｄの距離以内に、カテゴリの異なる新規教示点が入った場合である。仮に、既存の教示点がカテゴリＰに属し、新規教示点がカテゴリＱに属するとする。この場合、新規教示用データ入力以前は既存の教師点のＴＨｄの距離以内の近傍に対応する分類対象点Ｃは全てカテゴリＰと分類されるが、新規教示用データ入力以降は分類対象の最近傍教示点が既存の教示点になるか、新規教示点になるかでカテゴリはＰあるいはＱになる空間領域が特徴量ベクトル空間内にできる。このようなケースは前記のように、新規教示点が既存の教示点のＴＨｄの２倍の距離以内に置かれたとき発生する。しかしながら、異なるカテゴリを持つ教示用データ点が近傍にあること自体が、本来、カテゴリの異なる欠陥を分離すべく設計された特徴量ベクトル空間の前提と矛盾しているのであるから、既存の教示点の２×ＴＨｄなる距離以内にカテゴリの異なる新規教示点が置かれたときには、新規教示データに対応する入力欠陥画像及び既存の教示データに対応する入力欠陥画像を図２に示すモニタ３８上に並置して表示し、オペレータに教示データの入力に矛盾がないか問い合わせを行なう。オペレータは、モニタ３８上に並置された画像とそのカテゴリを確認し、どうちらか一方のカテゴリに統一、異なるカテゴリとしてそのまま登録、登録取りやめのいずれかの処理を選択する。これにより矛盾したデータがオペレータに感知されずに入力されることを防ぐことが出来る。このような処理を起動するか否かを決定するしきい値をここでは２×ＴＨｄとしたが、余裕を見ることができるようにα×ＴＨｄとし、αはオペレータが任意の値に設定できるよう外部入力可とすることも考えられる。
【００５６】
ここで述べた、いわゆる登録教示用データの矛盾は、分類過程と並行して実施される学習過程のみで起こる問題ではなく、初期の学習過程でも起こりうる。例えば、操作者（オペレータ）の入力ミス、未熟な操作者によるカテゴリの一貫性を欠いた教示用データ入力などが挙げられる。しかし初期の学習過程においても前述と同様なチェックを行なうことにより、このような誤入力を未然に防ぐことが可能となる。
以上により分類処理を行いながら、教示用データを追加していくことが出来るので、半導体ウエハ製造プロセスの変動により生起する欠陥分類属性（カテゴリ）と欠陥画像の対応の変化に、欠陥分類装置を追従させることが可能となる。また、プロセスの変動によって生じる未知の新たな欠陥を既に学習した欠陥属性（カテゴリ）から分別して、ユーザに新規の欠陥発生を通達可能な半導体欠陥の画像分類方法及び装置を提供が可能となる。
以上、特徴量の第１の実施例として、欠陥の色情報、形状、サイズなど予め決められたものとして説明を進めた。これらの特徴量の第１の実施例によれば、色情報、形状、サイズなどであるため、操作者に対してその値の意味が明示的であり、教示用データが適切に特徴量ベクトル空間内に分布しているか、分類対象欠陥から抽出される特徴量が妥当かなどが直感的に把握可能であるという利点がある。特に、光学画像の場合、色情報、形状、サイズなど特徴量を抽出することによって、カテゴリが付与できるように識別可能となる。
【００５７】
他方、電子線を照射して得られる二次電子や反射電子による画像の場合、色情報は得られず、その代わりにテクスチャーの状態（濃淡値の微妙な変化）、複雑な形状などの明示的でない特徴量でカテゴリが付与できるように識別する必要が生じる。しかしながら、テクスチャーの状態、複雑な形状などの特徴量の場合、明示的に設定することが困難で、そのため取り扱わなければならないカテゴリを十分に識別できる保証がない。即ち、テクスチャーの状態、複雑な形状などの明示的でない教示用欠陥画像をモニタ３８の表示領域４５に表示させても、オペレータはどのカテゴリＡを付与すればよいか容易に決定することができない。
【００５８】
そこで、この課題を解決することができる特徴量の第２の実施例について説明する。
光学画像でも、電子線画像でも、欠陥番号ｋの教示用画像信号としては、ラスター走査することによって、走査順に得られる濃淡値で示される画素信号（ｉ１、ｉ２，…，ｉｐ）で表されることになる。
ところで、このような教示用画像信号から撮像条件等による感度のばらつきの影響を受けない特徴量を算出するために、まず撮像条件等による感度のばらつきを除去した正規化された教示用画像信号を得る必要がある。そこで、画像処理装置３５は、画像記録装置３４で収集された欠陥番号ｋの教示用画像信号に対して、まず、２次元の大きさ（ｐ＝ｕ×ｗ）および明るさ（濃淡値）について正規化処理を行う。その結果、図１４に示すように予め一定の２次元の大きさ（ｐ＝ｕ×ｗ）に正規化され、更に明るさ（濃淡値）についても正規化された、ｕ×ｗの２次元の画素の濃淡値で構成されるベクトルで表される欠陥番号ｋの教示用画像信号Ｉｋ（次の（数９）式で示す。）が得られる。
【００５９】
【数９】

【００６０】
ｔは行列の転置を表す記号である。ｐは画素数である。
教示用画像信号Ｉｋは、上記のように、［ｉ１、ｉ２，…，ｉｐ］の転置行列に対して、予め一定の大きさ（ｐ＝ｕ×ｗ）に正規化し、画像の明るさ（濃淡値）についても上記（数９）式の右の項（濃淡値の平方和）に示すように正規化を施したものであるため、撮像条件等による感度のばらつきを除去した正規化された教示用画像信号を得ることができる。
そこで、特徴量の第２の実施例は、教示用欠陥画像について、共分散行列の固有ベクトルが張る固有空間（固有ベクトル空間）で表すものである。教示用欠陥画像について、共分散行列の固有ベクトルが張る固有空間を計算するために、まず、画像処理装置３５は、画像記録装置３４で収集された全ての正規化された教示用欠陥画像の平均画像Ａを次に示す（数１０）式によって求める。
【００６１】
【数１０】

【００６２】
次に、画像処理装置３５は、各教示用欠陥画像（Ｉ１，Ｉ２，…，Ｉｎ）から平均画像Ａを差し引き、行列Ｘを次に示す（数１１）に基づいて生成する。
Ｘ≡［Ｉ１−Ａ，Ｉ２−Ａ，…、Ｉｎ−Ａ］ （数１１）
そして、画像処理装置３５は、教示用欠陥画像集合の共分散行列Ｑを次に示す（数１２）式に基づいて計算する。
Ｑ≡Ｘ・（Ｘの転置行列） （数１２）
ｐ次元の固有ベクトル［ｅ1，ｅ2，…，ｅs］で張られる固有空間（部分空間）は次に示す固有方程式（数１３）を解くことによって求められる。
【００６３】
λｑ・ｅｑ＝Ｑ・ｅｑ （数１３）
即ち、教示用欠陥画像の特徴量を、ｓ個の大きい固有値（λ１≧…≧λｓ≧…≧λｐ）に対応する固有ベクトル［ｅ1，ｅ2，…，ｅs］で張られる固有空間（部分空間）で表現することができる。なお、ｓはｐ（＝ｕ×ｗ）より小さい値である。
ｓの値を大きく取ればとるだけ、固有ベクトル［ｅ1，ｅ2，…，ｅs］によって教示用欠陥画像の細部の特徴量を表現でき、ｓの値が小さければ固有ベクトル［ｅ1，ｅ2，…，ｅs］によって大まかな教示用欠陥画像の特徴量のみが表現できる性質を持っている。固有ベクトル［ｅ1，ｅ2，…，ｅs］で張られる固有空間によって、教示用欠陥画像の特徴量を表現することができる。
もともとの教示用欠陥画像と、次元ｓの固有空間で表現された画像との残差Ｅは、次に示す（数１４）式で表され、残差率Ｒｅは、次に示す（数１５）式で表される。
【００６４】
【数１４】

【００６５】
【数１５】

【００６６】
従って、ホストコンピュータ４０は、画像処理装置３５に対して提供する固有値（λ１≧…≧λｓ≧…≧λｐ）によって算出される残差率Ｒｅを指標として、分類正解率などの分類性能を勘案してｓを決めて画像処理装置３５に対して提供する。
以上説明したように、第２の実施例における特徴量とは、ｐ次元の固有ベクトル［ｅ1，ｅ2，…，ｅs］である。即ち、画像処理装置３５は、固有ベクトル［ｅ1，ｅ2，…，ｅs］を算出してホストコンピュータ４０に提供することになる。なお、固有ベクトルの算出も、ホストコンピュータ４０において実行してもよい。
【００６７】
ホストコンピュータ４０は、画像処理装置３５から提供を受けたまたは自身で求めた固有ベクトルを用いて、欠陥番号ｋの教示用欠陥画像について、次に示す（数１６）式に基づく固有ベクトル空間内の教示点ｉの位置にプロットすることができる。図１５には、一つの固有ベクトク［ｅ1］空間内における教示点ｉの位置を模式的に示した。なお、この際、ホストコンピュータ４０は、（Ｉｋ−Ａ）に関するデータを画像処理装置３５または画像記録装置３４から提供を受ける必要がある。
【００６８】
ｉｋ＝（［ｅ1，ｅ2，…，ｅs］の転置行列）×（Ｉｋ−Ａ） （数１６）
そして、ホストコンピュータ４０は、ベクトルｉｋのｓ個の要素と欠陥番号ｋのカテゴリを図４（ｃ）に示す教示用データの内容として記憶装置４３に記憶する。また、併せて固有ベクトル［ｅ1，ｅ2，…，ｅs］（第２の実施例における特徴量の座標軸となる。）も特徴量として記憶装置４３に記憶する。
以上の説明が、特徴量の第２の実施例に基づく図３（ａ）に示す特徴量抽出ステップＳ３３である。なお、教示データ作成ステップおよび学習ステップＳ３４については、特徴量の第１の実施例と同様に、即ち図６〜図９に示すように実行される。即ち、特徴量ベクトルが、特徴量の第１の実施例の場合、［ｃ，ｃ，…，ｅm］であるのに対して、特徴量の第２の実施例の場合、固有ベクトル［ｅ1，ｅ2，…，ｅs］となる。
次に、この特徴量の第２の実施例を用いた分類過程について説明する。画像記録装置３４から入力した分類対象欠陥画像を、画像処置装置３５において、教示用欠陥画像と同様の大きさに正規化し、明るさ（濃淡値）についても同様の正規化を施す。正規化処理後の分類対象欠陥画像ベクトルをＪとする。次に、画像処理装置３５またはホストコンピュータ４０において、このベクトルＪを次に示す（数１７）式によりｐ次元の固有ベクトル空間上の点ｊに投影する。
【００６９】
ｊ＝（［ｅ1，ｅ2，…，ｅk］の転置行列）×（Ｊ−Ａ） （数１７）
図１１に示すステップＳ１０２におけるカテゴリ決定は、この分類対象点ｊが固有ベクトク空間内に点在する教示点ｉｋのどの点に最も近いかで実行される。分類対象点の近傍にある教示点の探索とそれによるカテゴリ決定は、（数６）式と同様にして、次に示す（数１８）式に基づいて処理される。
【００７０】
【数１８】

【００７１】
｜ｉｋｊ｜は教示点座標ｉｋと分類対象点座標ｊのユークリッド距離を表す。分類未知欠陥の処理は前述した内容と同じである。分類過程において並行して実行される学習過程においては、新規教示用データも加えて固有ベクトル［ｅ1，ｅ2，…，ｅs］を計算し直した方が良い。固有ベクトルを計算し直しても、ｓ次元の空間に教示点は点在していること、教示点に対応するカテゴリは不変なこから、学習結果に継続性は保証される。
【００７２】
以上、特徴量の第２の実施例を用いても分類対象の欠陥画像に対して正確に分類カテゴリＢを付与することが可能となる。特に、波長の短いエキシマレーザ光などの紫外線による光学画像や、電子線を照射したことによって得られる粒子線画像等においては、特徴量として色情報が得られず、特徴量としてテクスチャーの状態（濃淡値の微妙な変化）、複雑な形状などが得られることから、特徴量ベクトル空間として第２の実施例である固有ベクトル空間を用いることが好ましい。
【００７３】
【発明の効果】
本発明によれば、欠陥画像の分類処理を行いながら、教示用データを追加していくことが出来るので、半導体ウエハ製造プロセスの変動により生起する欠陥分類属性と欠陥画像の対応の変化に、欠陥分類を追従させることができる効果を奏する。
また、本発明によれば、プロセスの変動によって生じる未知の新たな欠陥を既に学習した欠陥属性から分別して、ユーザに新規の欠陥発生を通達可能にすることができ、その結果、条件変動する製造プロセスで発生する新規の欠陥を確実に検出することが可能となるので、信頼性の高いプロセス管理のための欠陥情報収集を実現することができる効果を奏する。
【図面の簡単な説明】
【図１】本発明に係る欠陥分類方法およびその装置の半導体製造プロセスにける役割を説明するための図である。
【図２】本発明に係る欠陥分類装置の一実施例を示す構成図である。
【図３】本発明に係る欠陥分類装置が行う学習過程と分類過程との夫々の処理手順を説明するための図である。
【図４】本発明に係る学習過程における教示用データ作成手順、モニタ画面および教示用データを示す図である。
【図５】本発明に係る学習の原理を説明するための図である。
【図６】本発明に係る教示用データ作成手順を具体的に示すフローチャート図である。
【図７】図６に示す教示用データ作成手順を実行するためのモニタ画面を示す図である。
【図８】本発明に係る教示用データであるカテゴリ一覧表の一例を示す図である。
【図９】図６に示す教示用データ作成手順を実行するためのモニタ画面を示す図である。
【図１０】本発明に係る教示用欠陥画像における２つの特徴量に基づく特徴量ベクトル空間の第１の実施例における入力された教示点座標と既入力教示点座標との関係を模式的に示した図である。
【図１１】本発明に係る分類処理手順を具体的に示すフローチャート図である。
【図１２】分類対象欠陥画像に対して分類カテゴリを付与する際、特徴量ベクトル空間における分類対象点の近傍に存在する教示点の教示用欠陥画像をモニタの画面に出力して表示した図である。
【図１３】本発明に係る分類処理手順を実行するためのモニタ画面を示す図である。
【図１４】本発明に係るラスター走査によって得られる濃淡値の教示用欠陥画像を示す図である。
【図１５】本発明に係る特徴量ベクトル空間の第２の実施例における固有ベクトル空間における教示点を示す図である。
【符号の説明】
１…プローブ検査工程、２…外観検査装置、３…欠陥分類装置、４…外観検査結果、５…分類対象欠陥画像、６…不良モード別の発生頻度を示す画面、７…品質管理システム、８ａ、８ｂ…基板、２７…分類パラメータ（教示用データ）、２８…基板搬送装置、２９…光学系、３０…分類データ、３１…ステージ、３２…撮像装置、３３…画像入力装置、３４…画像記録装置（画像記憶装置）、３５…画像処理装置、３６…ステージ制御部、３７…基板搬送制御部、３８…モニタ、４０…ホストコンピュータ、４１…入力手段、４２…ネットワーク、４３…記憶装置、４５…入力欠陥画像を表示する表示領域、４７…入力カテゴリを表示するカラム、５２…欠陥画像詳細情報欄、５７…近傍既入力欠陥画像を表示する表示領域、５８…カラム、５９…ボタン、６２…カテゴリ一覧表示ボタン、６６…分類未知欠陥画像等の表示領域、６７…教示用欠陥画像等の表示領域、７３…カテゴリ入力カラム、７４…欠陥画像詳細情報欄、７５…カテゴリ表示カラム。[0001]
BACKGROUND OF THE INVENTION
BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for classifying an image based on contents included in the image, and more particularly to a defect image classification method and apparatus suitable for efficiently reviewing defects in a semiconductor manufacturing process, and for teaching. It relates to a data creation method.
[0002]
[Prior art]
Conventionally, for example, as a method of classifying defect attributes of a semiconductor wafer pattern based on an image thereof, a method using a neural network is known as described in JP-A-8-21803. In this method, first, image data called “teaching image data” is prepared for each defect attribute, and learning of the neural network is executed using the image data. In the learning, the error value obtained by inputting the teaching image a number of times until the error value indicating the discrepancy between the input teaching image and the classification result that is the output of the neural network converges to a certain value or less. It is executed by repeating the procedure of reduction according to the procedure in the weighting factor of the network. The weighting coefficient data of the neural network obtained as a result of learning is stored as learning data. When performing classification, the learning attribute is used to classify the defect attributes of the input defect image.
[0003]
[Problems to be solved by the invention]
When semiconductor wafer pattern defects in the semiconductor wafer manufacturing process are classified using the above-mentioned conventional classification method, learning is performed before classification execution. However, only the defect attribute obtained at the time of learning can be recorded in the learning data. For this reason, even if a defect attribute and the attribute defined by the user are the same due to the variation in the semiconductor wafer manufacturing process that occurs during the classification, the user is informed that the defect shape and pattern are different from those at the time of learning. It cannot follow to classify into the attribute indicated by. In addition, when the newly acquired defect image is included in the teacher image data and re-learning is performed, the conventional classification performance is not guaranteed due to the change of the learning data. On the other hand, the neural network has a generalization ability that can recognize and execute classification of attributes in the light of similarity to a pattern obtained at the time of learning even for inputs that are not presented at the time of learning. However, this generalization capability is not sufficient to follow the process fluctuations described above, and conversely, unknown new defects caused by process fluctuations are classified into known defect attributes acquired during learning, and the process There is a problem such as having a risk of missing an unknown defect which is an important clue for extracting new problems.
[0004]
An object of the present invention is to solve the above-mentioned problems on a semiconductor wafer or the like that can follow a change in correspondence between a defect classification attribute and a defect image caused by a change in a manufacturing process of the semiconductor wafer or the like that occurs during classification. It is an object of the present invention to provide a method and a device for classifying defects generated in the above, and a method for creating teaching data.
Another object of the present invention is to classify an unknown new defect caused by process variation from a defect attribute that has already been learned, and to provide a user with a new defect attribute generation method and apparatus for classifying the defect. Another object of the present invention is to provide a teaching data creation method.
[0005]
[Means for Solving the Problems]
In order to achieve the above object, the present invention provides a plurality of teaching defect images and a category corresponding to a defect type by giving a category corresponding to a defect type to a plurality of teaching defect images in advance. A learning process for learning and acquiring teaching data indicating the correspondence relationship between and a classification target defect image obtained by imaging the inspection target object based on the teaching data acquired in the learning process A classification process for classifying defect types by assigning classification categories, and when a new classification category is assigned to a classification target defect image in the classification process, the new classification category and the new classification category are A defect classification method characterized by comprising an update process for updating teaching data acquired in the learning process based on a given classification target defect image.
Further, the present invention is characterized in that the update of the teaching data is a write-once type in the update process of the defect classification method.
Further, the present invention provides a teaching that shows a correspondence relationship between a plurality of defect images for teaching and a category corresponding to the type of defect by giving a category corresponding to the type of defect to a plurality of defect images for teaching in advance. A learning process for learning and acquiring data, and a classification target defect image obtained by imaging an object to be inspected is classified into known classification categories based on the teaching data acquired in the learning process. If it is possible to classify, and if classification is possible, a classification process for classifying the defect type by assigning a known classification category to the classification target defect image, and classification is not possible in the classification process In this case, the defect classification method is characterized by having a presentation process of presenting a classification target defect image relating to a classification unknown defect that cannot be classified.
[0006]
Further, the present invention is characterized in that, in the classification process of the defect classification method, when classification is impossible, information relating to classification unknown defects that cannot be classified is stored.
[0007]
Further, the present invention is characterized in that in the classification process of the defect classification method, information on the classification target defect to which a known classification category is assigned is displayed on the display means. The present invention is characterized in that, in the process of presenting the defect classification method, a new classification category is assigned to the classification target defect image of the presented classification unknown defect. According to the present invention, in the presentation process of the defect classification method, a teaching defect image having a high degree of similarity is obtained from the teaching data acquired in the learning process with respect to the classification target defect image of the presented classification unknown defect. It is characterized by presenting.
Further, the present invention provides a teaching that shows a correspondence relationship between a plurality of defect images for teaching and a category corresponding to the type of defect by giving a category corresponding to the type of defect to a plurality of defect images for teaching in advance. A learning defect image having a high degree of similarity from the teaching data acquired in the learning process, with respect to a learning process in which the learning data is acquired and the classification target defect image obtained by imaging the inspection object And a classification process of classifying the defect type by assigning a classification category to the classification target defect image to a category source given to the searched defect image with high similarity. This is a method for classifying defects.
Further, the present invention provides a teaching that shows a correspondence relationship between a plurality of defect images for teaching and a category corresponding to the type of defect by giving a category corresponding to the type of defect to a plurality of defect images for teaching in advance. A learning defect image having a high degree of similarity from the teaching data acquired in the learning process, with respect to a learning process in which the learning data is acquired and the classification target defect image obtained by imaging the inspection object Are searched for, and the retrieved defect image for teaching with high similarity and the category assigned to the teaching defect image with high similarity are presented, and the displayed teaching defect image with high similarity and A defect classification method comprising: classifying a defect type by assigning a classification category to the classification target defect image based on a category assigned to a teaching defect image having a high similarity is there.
[0008]
The present invention also provides a teaching data creation method for classifying types of defects based on a classification target defect image obtained by imaging an object to be inspected, for a plurality of teaching defect images. A teaching data is created by assigning a category corresponding to the type of defect, a plurality of feature amounts are calculated for each of the teaching defect images, and each teaching amount in the calculated multidimensional feature amount vector space is calculated. Based on the Euclidean distance between the teaching point coordinates for the defect image, the teaching data comprising the correspondence between the teaching defect image and the category is diagnosed, and when it is necessary to correct by the diagnosis, the teaching data is corrected. It is characterized by doing.
The present invention also provides a teaching data creation method for classifying types of defects based on a classification target defect image obtained by imaging an object to be inspected, for a plurality of teaching defect images. A teaching data is created by assigning a category corresponding to the type of defect, a plurality of feature amounts are calculated for each of the teaching defect images, and each teaching amount in the calculated multidimensional feature amount vector space is calculated. By examining whether or not the Euclidean distance between the teaching point coordinates for the defect image satisfies a desired threshold value, the teaching data consisting of the correspondence between the teaching defect image and the category is diagnosed and corrected by the diagnosis If necessary, the teaching data is corrected.
[0009]
Further, the present invention creates teaching data by assigning a category corresponding to the type of defect to a plurality of defect images for teaching, calculates a plurality of feature amounts for each of the teaching defect images, Based on the Euclidean distance between the teaching point coordinates for each teaching defect image in the calculated multidimensional feature vector space, the teaching data comprising the correspondence between the teaching defect image and the category is diagnosed, and the diagnosis If there is a need for correction by the learning process of correcting the teaching data, the teaching data acquired in the learning process is applied to the classification target defect image obtained by imaging the inspection target object. It is determined whether or not classification into a known classification category is possible, and if classification is possible, the type of defect is determined by assigning a known classification category to the classification target defect image. Classified, if not possible classification is a classification method of defects and having a classification process of storing the information on the classification unknown defect impossible classification.
Further, the present invention creates teaching data by assigning a category corresponding to the type of defect to a plurality of defect images for teaching, calculates a plurality of feature amounts for each of the teaching defect images, Based on the Euclidean distance between the teaching point coordinates for each teaching defect image in the calculated multidimensional feature vector space, the teaching data comprising the correspondence between the teaching defect image and the category is diagnosed, and the diagnosis When it is necessary to correct by the above, the learning process for correcting the teaching data, a plurality of feature amounts are calculated for the classification target defect image obtained by imaging the inspection target object, and the calculated Each item in the multidimensional feature vector space based on the target coordinate of the classification target defect image in the multidimensional feature vector space and the teaching data acquired in the learning process Based on the Euclidean distance between the teaching point coordinates of the display defect image, it is determined whether or not the image can be classified into a known classification category. A defect classification method comprising: classifying the defect type by assigning a category, and storing the defect type as information relating to an unclassified unidentified defect when the defect type cannot be classified. .
[0010]
Further, the present invention is characterized in that in the defect classification method, information on unknown defects obtained in the classification process is fed back to the learning process to update the teaching data.
In the classification process of the defect classification method, the present invention outputs the number of defects for each classified defect type.
Further, the present invention is characterized in that the defect classification method further includes a defect generation cause estimation step of estimating a defect generation cause based on a defect type classified in the classification step.
[0011]
In addition, the present invention provides a defect image detection unit that images a plurality of teaching defects in advance and detects a plurality of teaching defect images, and a plurality of teaching defect images detected by the defect image detection unit. A display unit that displays the category corresponding to the type of the defect and the plurality of teaching defects in association with each other, and indicates a correspondence relationship between the plurality of teaching defects and the category corresponding to the type of defect A learning unit that obtains data for storage and stores it in a storage unit, a defect image detection unit that captures an object to be inspected and detects a defect image to be classified, and a classification target defect image detected by the defect image detection unit On the other hand, in a classification unit including a calculation unit that classifies the defect type by assigning a classification category based on teaching data stored in the storage unit of the learning unit, and in the calculation unit of the classification unit Missing classification target When a new classification category is assigned to the image, the teaching data acquired by the learning unit is updated based on the new classification category and the classification target defect image to which the new classification category is assigned. It is a defect classification device characterized by comprising update means.
In addition, the present invention provides a defect image detection unit that images a plurality of teaching defects in advance and detects a plurality of teaching defect images, and a plurality of teaching defect images detected by the defect image detection unit. A display unit that displays the category corresponding to the type of the defect and the plurality of teaching defects in association with each other, and indicates a correspondence relationship between the plurality of teaching defects and the category corresponding to the type of defect A learning unit that obtains data for storage and stores it in a storage unit, a defect image detection unit that captures an object to be inspected and detects a defect image to be classified, and a classification target defect image detected by the defect image detection unit On the other hand, calculation means for determining whether or not classification into a known classification category is possible based on the teaching data stored in the storage means of the learning unit, and if the calculation means determines that classification is not possible, Impossible classification unknown A classification device of the defect, characterized in that it consists of a classification unit that includes a presentation means for presenting information about Recessed.
[0012]
Further, the present invention is characterized in that, in the classification unit of the defect classification device, the calculation means creates a reference list for classification target defect images of classification unknown defects determined to be unclassifiable. And
In addition, the present invention includes a display unit that displays information about a classification target defect to which a known classification category is assigned when the classification unit of the defect classification apparatus determines that the calculation unit can be classified. It is characterized by.
According to the present invention, the classification unit of the defect classification apparatus further includes means for inputting a classification category for the classification target image presented by the presentation means.
[0013]
In addition, the present invention provides a defect image detection unit that images a plurality of teaching defects in advance and detects a plurality of teaching defect images, and a plurality of teaching defect images detected by the defect image detection unit. A display unit that displays the category corresponding to the type of the defect and the plurality of teaching defects in association with each other, and indicates a correspondence relationship between the plurality of teaching defects and the category corresponding to the type of defect A learning unit that obtains data for storage and stores it in a storage unit, a defect image detection unit that captures an object to be inspected and detects a defect image to be classified, and a classification target defect image detected by the defect image detection unit On the other hand, with respect to the teaching defect image having a high similarity from the teaching data stored in the storage means of the learning unit, and the teaching defect image having a high similarity searched by the calculating means Type of defect granted And a presenting means for presenting the corresponding category is a classification device of the defect, characterized in that it consists of a structure with classification unit to classify the type of defect with respect to the classification target defect image.
In addition, the present invention provides a defect image detection unit that images a plurality of teaching defects in advance and detects a plurality of teaching defect images, and a plurality of teaching defect images detected by the defect image detection unit. A display unit that displays the category corresponding to the type of the defect and the plurality of teaching defects in association with each other, and indicates a correspondence relationship between the plurality of teaching defects and the category corresponding to the type of defect A learning unit that obtains data for storage and stores it in a storage unit, a defect image detection unit that captures an object to be inspected and detects a defect image to be classified, and a classification target defect image detected by the defect image detection unit On the other hand, a calculation unit that searches for a teaching defect image having a high similarity from the teaching data stored in the storage unit of the learning unit, and a teaching defect image that has a high similarity searched by the calculation unit and the similarity High-defective teaching image And a classifying unit configured to classify the defect type for the defect image to be classified, with a presentation unit that presents a category corresponding to the type of defect assigned to the defect classification Device.
[0014]
Further, according to the present invention, in the defect classification apparatus, when the classification unit assigns a classification category different from the teaching data to the classification target defect image, the teaching data is fed back to the learning unit to update the teaching data. It is characterized by comprising.
In addition, the present invention provides a defect image detection unit that images a plurality of teaching defects in advance and detects a plurality of teaching defect images, and a plurality of each of the plurality of teaching defect images detected by the defect image detection unit. For teaching, comprising correspondence between the teaching defect image and the category based on the Euclidean distance between the teaching point coordinates for each teaching defect image in the calculated multidimensional feature amount vector space A learning unit for acquiring teaching data indicating a correspondence relationship between a plurality of teaching defects and a category corresponding to a defect type and storing the teaching data in a storage unit, and a target to be inspected. A defect image detection unit that picks up an image and detects a classification target defect image, and a classification target defect image detected by the defect image detection unit, based on teaching data stored in the storage unit of the learning unit. A calculation means for determining whether or not classification is possible, and types of defects classified according to a known classification category assigned to the classification target defect image when the calculation means determines that classification is possible A defect classifying device comprising: a classifying unit comprising: storage means for storing information relating to, and storage means for storing information relating to unidentifiable defects that cannot be classified when it is determined that classification is impossible is there.
[0015]
In addition, the present invention provides a defect image detection unit that images a plurality of teaching defects in advance and detects a plurality of teaching defect images, and a plurality of each of the plurality of teaching defect images detected by the defect image detection unit. For teaching, comprising correspondence between the teaching defect image and the category based on the Euclidean distance between the teaching point coordinates for each teaching defect image in the calculated multidimensional feature amount vector space A learning unit for acquiring teaching data indicating a correspondence relationship between a plurality of teaching defects and a category corresponding to a defect type and storing the teaching data in a storage unit, and a target to be inspected. Defect image detection means for picking up an image to detect a classification target defect image, and calculating a plurality of feature quantities for the classification target defect image detected by the defect image detection means, and calculating the calculated multidimensional feature quantity vector Between the classification target point coordinates for the classification target defect image in the space and the teaching point coordinates for each teaching defect image in the multidimensional feature vector space based on the teaching data stored in the storage means of the learning unit A calculation means for calculating a Euclidean distance and determining whether or not the classification can be made into a known classification category based on the Euclidean distance between the calculated classification target point coordinates and the teaching point coordinates, and classification by the calculation means Information regarding the type of defect classified by the known classification category assigned to the classification target defect image, and information regarding an unclassified defect that cannot be classified when it is determined that classification is impossible. A defect classifying apparatus comprising: a classifying unit configured to include a storage unit for storing.
[0016]
The present invention also uses a plurality of preset feature amounts (reference feature amounts) or a plurality of feature amounts (reference feature amounts) extracted from a plurality of teaching image sets in order to extract image data features. Then, using a multidimensional feature vector space formed by this reference feature value, a teaching defect image is associated with one point in the feature value vector space as a teach point coordinate by the reference feature value, and a plurality of teachings are similarly performed. A defect image for use is recorded as a teaching point coordinate in the feature vector space, and the teaching point coordinate has position information in the feature vector space and also has a defect classification attribute (classification category) to which the corresponding image data belongs. In the same way, the learning process to be performed and the classification target defect image inputted as the classification target are classified into the feature vector space by the reference feature quantity (the point generated by the classification target image). And by association, then search the teaching points in the closest to the classified point, and a classification process of the defect classification attribute (classification category) a defect attribute of the classification target holding the teaching point. However, if the relationship between the classification target point and the searched teaching point is not determined to be near in light of a predetermined distance measure, the classification target point is classified as an unknown attribute, that is, an unknown defect. When the defect classification attribute is presented to the unknown defect by the operator, the presented defect classification attribute is attached to the unknown defect. Furthermore, by additionally registering the classification target points as teaching data in the teaching points, learning can be executed in parallel with the classification work. Thereby, it is possible to follow the change of the correspondence between the defect classification attribute and the defect image caused by the variation of the semiconductor wafer manufacturing process, and the unknown new defect caused by the variation of the process is separated from the already learned defect attribute, The user can be notified of the occurrence of new defect attributes.
[0017]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a diagram for explaining a role played by a defect classification method and apparatus according to the present invention in a semiconductor manufacturing process. A semiconductor manufacturing process is manufactured through several hundred processes 1 to n, and may take a period of about 100 days to complete. However, it is the probe inspection 1 in which all processes are completed to determine the quality of the product. For this reason, in order to improve the yield, a means for estimating the quality of an intermediate process is essential. For this purpose, for example, an appearance inspection by the appearance inspection apparatus 2 is performed on a substrate (wafer) 8 processed in a certain manufacturing process, and the quality of the process is determined from the appearance abnormality of the wiring pattern. ing. When a process abnormality is confirmed, it is necessary to take countermeasures. However, the defect classification apparatus 3 plays an important role as information collection means for this purpose. That is, from the result 4 of the appearance inspection, the location and number of defects can be grasped, but information such as the nature of the defects cannot be obtained.
For this reason, the defect classification device 3 captures the image 5 at the defect site (defect location) obtained from the appearance inspection device 2, and based on this, the type of defect (whether it is a foreign matter defect or a wiring pattern defect) The quality control system 7 displays the occurrence frequency 6 for each failure mode on the basis of the classified information using the display means, so that the operator can The countermeasure candidates can be narrowed down in the process 9a for narrowing down countermeasure candidates. By the way, manufacturing process steps include various film forming steps, resist coating steps, exposure / development steps, etching steps, resist stripping steps, and steps of planarizing by chemically and mechanically polishing the surface of the insulating film. It is comprised by the manufacturing line by which the various process processing apparatuses were comprised. Therefore, there are various process processing apparatuses that generate various types (categories) of defects.
[0018]
When the occurrence frequency 6 for each failure mode is displayed in the quality management system 7, it can be seen that in the example of FIG. 1, the occurrence frequency of foreign matter is larger than that of the pattern defect, so that it is sufficient to take measures to prevent the occurrence of foreign matter. In other words, the worker preferentially estimates the cause of defects having a high occurrence frequency in the cause estimation process 9b, and measures the process processing apparatus that executes process processing based on the estimated cause in the countermeasure process 9c. By applying this, the yield can be improved quickly.
As described above, the defect classification apparatus 3 according to the present invention is an apparatus that automates the defect classification work that has been conventionally performed visually, and requires high classification performance.
[0019]
Next, an embodiment of the defect classification method and apparatus according to the present invention will be described with reference to FIGS. FIG. 2 is a block diagram showing an embodiment of the defect classification apparatus according to the present invention. FIG. 3 is a diagram for explaining a classification procedure performed by the defect classification apparatus shown in FIG. FIG. 4 shows an example of a teaching data creation procedure, a screen, and teaching data. FIG. 5 is a diagram for explaining the principle of learning. The description will be given with reference to FIG. 2 according to the classification procedure of FIG. Reference is made to FIGS. 4 and 5 where applicable.
The defect classification procedure is divided into a learning process shown in FIG. 3A and a classification process shown in FIG.
[0020]
In the learning process, first, a defect image collection step (process) S31 for creating teaching data is performed. That is, in step S31, defect images for creating teaching data are collected and stored in the image recording device (image storage device) 34. For this reason, first, the substrate 8a on which a defect for teaching selected as appropriate is formed is mounted on the stage 31 by the substrate transfer device 28 shown in FIG. On the other hand, the host computer 40 receives defect coordinate information corresponding to the teaching substrate 8a selected by the visual inspection apparatus 2 or the like as the defect in the teaching data from the host system via the network 42. The host system refers to a quality management system such as a yield management system, a production line monitoring system, a process management system, or an inspection apparatus 2 for inspecting defects. The host computer 40 refers to the defect coordinate information, sends an instruction to the stage control unit 36, moves the stage 31, and moves the teaching defect on the substrate 8a to the observation position. An imaging device 32 such as a TV camera captures a defect in teaching data obtained via the optical system 29, and an image recording device (image memory) 34 captures the captured teaching defective image via the image input device 33. To record (memorize). By repeating the above processing for the designated teaching defect, the teaching defect image signal is accumulated in the image storage device 34 and collected. As the teaching defect image, a wafer 8a having foreign matters, wiring pattern defects, scratches, and other defects generated in various categories in the semiconductor manufacturing process may be appropriately selected. As described above, the substrate 8a is used for preparing teaching data in advance.
[0021]
Next, a teaching data creation step (process) S32 is executed by the operator. That is, in step S32, teaching data creation (category A addition) is performed by an operator's operation on the host computer 40. 4A is a flowchart for creating teaching data, FIG. 4B is a screen of the monitor 38 for creating teaching data, and FIG. 4C is an example of data contents of defect data recorded as teaching data in the storage device 43. is there.
[0022]
As shown in FIG. 4A, in step S41, the operator (operator) designates the defect image i = 1 using the information input means 41 such as a keyboard or a mouse. Then, the defective image signal i = 1 for teaching stored in the image recording device (image storage device) 34 is called and displayed on the monitor 38. At the same time, in the process of sequentially observing the defect images for teaching, the operator registers the similar defect images as the same category, and the category and defect name (defect type: foreign matter, pattern defect, scratch, etc.) Is input to the host computer 40 using the input means 41 and registered in the storage device 43. As a result, the host computer 40 can display the category list 32 indicating the correspondence between the category and the defect name on the screen of the monitor 38. In step S43, as shown in FIG. 4B, the operator observes the teaching defect image i = 1 displayed as 31 on the screen of the monitor 38, and displays a category list (correspondence between categories and defect names). ) 32, the category number (33 on the screen of the monitor 38) is input and assigned to the teaching defect image i = 1 by the input means 41, so that the defect number i = 1 indicating the teaching defect image is given. Category A is stored in the storage device 43. In step S <b> 44, until the processing is completed for all the teaching defects collected and recorded in the image recording device 34, the operator repeats the designation operation of the teaching defect image i = i + 1 using the input unit 41 in step S <b> 45. In step S42, the host computer 40 repeats sequentially calling the teaching defect image signal i = i + 1 recorded in the image recording device 34 and displaying it on the screen of the monitor 38. Further, in step S43, the operator refers to the category list (correspondence between category and defect name) 32 sequentially displayed on the monitor 38, and sequentially uses the input means 41 for the teaching defect image i = i + 1. By inputting and assigning A, category A is sequentially stored in the storage device 43 for the defect number i = i + 1 indicating the teaching defect image. As described above, the process is repeated until the category A is assigned to all the collected defects.
[0023]
That is, as shown in FIG. 4B, the teaching defect image 31 and the category list 32 are displayed on the screen of the monitor 38, so that the operator can observe the category defect image 31 while observing the teaching defect image 31. A corresponding category number is selected from the list 32 and input using the category number input unit 33. Here, the category is a number given to the type of defect, and is determined according to the type of defect. For example, the foreign matter is 1 and the pattern defect is 2. The category input by the operator using the information input means 41 such as a keyboard or a mouse is stored in the storage device 43 as teaching data in the host computer 40. Here, it should be noted that the category is not determined in advance and is determined in the process of observing defects in order. That is, since the overall tendency of the collected defect images is unknown, it is necessary to register similar defects as the same category while sequentially observing the defect image i. For this reason, the teaching data tends to include contradictory data.
Next, in step S33, the image processing apparatus 35 extracts various feature amounts 35 (including selection of feature amounts) from each teaching defect image i and provides them to the host computer 40. Correspondingly, the data is stored in the storage device 43 as teaching data shown in FIG. That is, the image processing device 35 reads each teaching defect image signal i stored in the image recording device 34, performs image processing on each read teaching defect image signal i, and performs various feature amounts. Is extracted, provided to the host computer 40, and stored in the storage device 43. Here, the various feature amounts of the defect image include color information, gray value information, shape, size, and the like of the defect image. For example, the category 1 foreign material is dark and nearly circular, but the category 2 pattern defect has the same color and shape as the surrounding pattern, and the category 3 scratch has a small area but a long length. There is. Therefore, the image processing device 35 calculates predetermined types of feature amounts 1 to 5 (in the example of FIG. 4C, five ways 34) and provides them to the host computer 40, and stores them in the storage device 43 as teaching data. Stores corresponding to the defect number. Here, it should be noted that a feature quantity effective for classification cannot be uniquely determined. Defects differ depending on the product or process targeted by the defect classification device 3. For this reason, it is possible to prepare a large number of feature amounts (color information, gray value information, shape, size, etc.) within a conceivable range, and select and use only the feature amounts effective for the classification of collected defects. Done. A method for selecting the feature amount will be described later. Note that the feature extraction in step S33 shown in FIG.
[0024]
Next, in step S <b> 34, the host computer 40 performs learning (classification parameter) based on the teaching data to which category A is assigned in step S <b> 32, and stores the calculated classification parameter 27 in the storage device 43. 5A and 5B are diagrams for explaining the principle of learning. FIG. 5A is a diagram in which the teaching data to which category A is assigned in step S32 is plotted in the feature amount space. Here, the plot point 37 marked with x is category 1, the plot point 38 marked with □ is category 2, c1 is feature amount 1, and c2 is an axis indicating the value of feature amount 2. Learning performed in the host computer 40 means calculating a discriminant function between categories assigned in step S32, and this discriminant function (a parameter to be defined) is called a classification parameter, and is obtained as a result of learning. Is called learning data. The classification parameter is stored as learning data. When the discriminant function is illustrated, 41 shown in FIG. 5B is a boundary between categories. Many methods are known for calculating the discriminant function. In this specification, the discriminant function will be described by taking discriminant analysis as a representative method as an example. Hereinafter, the explanation will be made by taking discriminant analysis as an example, but the present invention can be similarly applied to other methods.
Next, discriminant analysis, which is a method for calculating a discriminant function between categories, which is the classification parameter 27, based on a plurality of feature amounts will be described with reference to FIG. In the figure, point 42 represents the center of gravity of category 1, point 43 represents the center of gravity of category 2, and ellipse 44 represents a collection of points having the same distance from the center of gravity. Here, the distance is a Mahalanobis distance obtained by standardizing the Euclidean distance in the feature space by the distribution of the teaching data. The boundary line 41 is a set of points that are equidistant from the centroids of category 1 and category 2. The discriminant function for the unknown point 45 (corresponding to the detected defect that performs defect classification) is given by the following equation (1).
Z_{1 ・ 2}= D₁ ²-D₂ ²                                        (Equation 1)
However, Z_{1 ・ 2}: Category 1 and Category 2 discriminant functions
D₁ ² : Mahalanobis square distance from the center of gravity of category 1
D₂ ² : Mahalanobis square distance from the center of gravity of category 2
For this reason, Z_{1 ・ 2}Classification is possible according to the following formula (Equation 2).
[0025]
Z_{1 ・ 2}<0 → belongs to category 1
Z_{1 ・ 2}> 0 → belongs to category 2 (Equation 2)
A neural network is positioned as a method for obtaining this discriminant function. Although FIGS. 5A and 5B have been described using two feature amounts, it should be noted that a multidimensional vector space is actually configured.
[0026]
Next, the inspection target substrate (wafer) 8 processed by the process processing apparatus actually installed in the desired process processing step according to the present invention is extracted in units of substrates or lots, and the appearance inspection apparatus 2 The classification process shown in FIG. 3B executed by the defect classification apparatus 3 on the substrate to be inspected in units of substrates or lots determined by the visual inspection apparatus 2 as a defect will be described.
In the classification process shown in FIG. 3B, in step S35, the substrate unit or lot unit for the inspected substrate (wafer) 8 actually processed by the process processing apparatus installed in the desired process processing step. The defect image of the substrate to be inspected in units of substrates or lots determined as defects by the appearance inspection device 2 is collected and image recording device (image storage device) 34. Stored in That is, first, as shown in FIG. 2, the substrate to be inspected 8 in units of lots or lots determined by the appearance inspection apparatus 2 by the substrate transfer device 28 controlled by the substrate transfer control unit 37 is placed on the stage 31. To be installed. On the other hand, the host computer 40 receives defect coordinate information corresponding to the substrate to be inspected 8 which is inspected in the substrate unit or lot unit by the appearance inspection apparatus 2 or the like and has a defect from the host system via the network 42. The host system refers to a quality management system such as a yield management system, a production line monitoring system, a process management system, or an inspection apparatus 2 for inspecting defects. The host computer 40 refers to the received defect coordinate information, sends an instruction to the stage control unit 36, moves the stage 31, and moves the defect on the inspection target substrate 8 to the observation position. An imaging device 32 such as a TV camera captures an image of a defect obtained through the optical system 29, and the image signal of the captured defect is input to an image recording device (image memory) 34 through an image input device 33. Record (memorize). By repeating the above processing for the defect inspected by the appearance inspection apparatus 2, a defect image signal for classifying the category is accumulated in the image storage device 34 in correspondence with the defect number to be classified, and the classification target A defect image signal will be collected.
[0027]
Next, in step S <b> 36, the image processing device 35 extracts various feature amounts (including feature amount selection) from the defect image to be classified, provides them to the host computer 40, and stores them in the storage device 43. . That is, the image processing device 35 reads out the classification target defect image signal stored in the image recording device 34 in correspondence with the classification target defect number, and performs image processing on each of the read defect image signals. Then, various feature amounts (referred to as classification data) of the defect image are extracted, provided to the host computer 40, and stored in the storage device 43. Here, the various feature amounts of the defect image include color information, gray value information, shape, size, and the like of the defect image. The classification data stored in the storage device 43 has the same contents as the teaching data shown in FIG. 4C, but the category column 36 is blank.
Therefore, the host computer 40 classifies the classification data (various feature quantities) for each defect stored in the storage device 43 and the classification parameters (features) calculated in the learning process and stored in the storage device 43 in step S37. A classification function for estimating a category (corresponding to the type of defect) using the teaching data indicated by the discriminant function between categories in the quantity space) 27 to estimate the classification category B30, and this estimated classification Category B (corresponding to the type of defect) 30 is stored in the storage device 43 in association with the defect number to be classified.
Thus, the stored classification category B30 is transferred to the quality management system 7 which is the host system via the network 42, and the failure mode (classification category B corresponding to the type of defect) as shown in FIG. It is displayed, and analysis such as narrowing down countermeasure candidates and estimating the cause of the defect is performed.
[0028]
In the learning type classification system (defect classification device) 3 as described above, since the learning is performed before the classification execution, the defect obtained at the time of learning is selected no matter how the teaching image data is examined and selected. Only attributes can be recorded as learning data.
For this reason, in fact, due to fluctuations in the semiconductor wafer manufacturing process that occur during classification execution, even if a certain defect has the same category defined by the user, the feature amount such as the defect shape differs from that during learning. In this case, it cannot be classified into the category designated by the user. That is, when the newly acquired defect image is included in the teaching image data and re-learning is executed, the classification parameter 27 changes. That is, the category boundary line 41 in FIG. Means that the classified performance is not guaranteed. In FIG. 5B, the boundary is a straight line. However, even if this is replaced with a higher-order curve, if the category boundary required as a result of relearning exceeds the order of the curve, the same situation occurs. According to the method using the discriminant analysis method such as classification using the neural network described above, this problem cannot be avoided as long as the order of the category boundary for discrimination is determined in advance and learning is performed.
Further, according to discriminant analysis, an unknown new defect caused by a process variation is classified into a known defect attribute (defect category) acquired at the time of learning. Although it is possible to control whether or not the classification is performed depending on the reliability of classification, it is difficult to determine a threshold value for reliability so that unknown defects are not classified, so that new problems in the process can be identified. There is no way to eliminate the risk of missing unknown defects that are important clues.
[0029]
Therefore, in the present invention, the classification process and the learning process can be performed in parallel so as to follow the change in the correspondence between the defect classification attribute and the defect image caused by the variation in the semiconductor wafer manufacturing process, and further the process variation The new defect attributed by the user is separated from the already learned defect attribute, so that the user can be notified of the generation of the new defect attribute.
[0030]
Next, a defect image classification method necessary for achieving this will be described. As described above, the classification of defect images can be broadly divided into a learning process and a classification process.
First, as various feature amounts, for example, geometric quantitative values such as the circularity and area of the defect portion, image average luminance values, dispersion values, and values relating to colors (R (red) G (green) B (blue ) Ratio, hue, distribution of color values, etc.), or numerical values related to image information including background images of defects.
[0031]
A first embodiment of defect image classification, which is roughly divided into a learning process and a classification process, based on this feature amount will be described.
[0032]
In the learning process, first, a teaching defect image is collected and recorded (stored) in the image recording device 34 in step S31. Next, in step S32, teaching data creation (category A addition) is performed by the operator's operation on the host computer 40. Next, in step S33, the image processing apparatus 35 extracts various feature amounts 35 (including selection of feature amounts) from each teaching defect image h and provides them to the host computer 40. Correspondingly, it is stored in the storage device 43 as teaching data. That is, in step S <b> 33, the image processing device 35 extracts m feature amounts set from the teaching defect image to be learned recorded in the image recording device 34. Now, let c be a feature quantity, and m be the number of preset feature quantity types. A set [m1, c2,..., Cm] of m feature quantities extracted from one teaching defect image is defined as a feature quantity vector. The feature vector Ch is a point in the feature vector space [c1_h, C2_h, ..., cm_h]. That is, the feature vector extracted for each teaching defect image in the image processing device 35 is provided to the host computer 40, and as shown in FIG. 4C, the defect number which is the index of each teaching image. Are stored in the storage device 43 together with the defect category. This recorded data is hereinafter referred to as learning data (teaching data). A point represented by a feature vector extracted from each defect image for teaching in the feature vector space is hereinafter referred to as a teaching point.
[0033]
Next, specific contents of the teaching data creation step S32, the feature amount extraction step S33, and the learning step S34 will be described with reference to FIGS.
That is, in step S61, by the operation of the input means 41 by the operator, the host computer 40 calls the first teaching defect image from the image recording device 34 and displays it on the display area 45 on the monitor 38 shown in FIG. The operator observes the displayed teaching defect image, determines in step S62 whether or not to register the teaching defect image, and if not registered, operates the scroll 55 in step S69 to operate the host computer 40. Calls the next defect image for teaching from the image recording device 34 and displays it on the display area 45 on the monitor 38. When the teaching defect image of the teaching defect number h is registered in step S62, the operator instructs the teaching data display 53 to display the category list shown in FIG. 8 in the display area 46, and in step S63. It is checked whether or not the category corresponding to the displayed teaching defect image is in the displayed category list. If there is, the operator selects a corresponding category number from the list and inputs the category number in step S64. The data is input to the unit 47 and a category is assigned. This process is performed sequentially for all defect images collected for teaching, and is repeated until the processing of all the defect images is completed.
[0034]
The contents displayed on the monitor 38 are shown in FIG. On the monitor 38, a teaching defect image is displayed in the display area 45 and a category list is displayed in the display area 46. A category is a code, number, or defect name itself given to a defect type, and is determined according to the defect type. FIG. 8 shows an example in which defect categories are defined by numbers, and defect names corresponding to the category numbers are displayed in the category list for easy understanding. If the defect category is expressed by a number, for example, 1 is set for a foreign object, and 2 is set for a pattern defect. In addition to the categories that can be assumed and set in advance, there are categories that are set and added in the process of sequentially observing defects.
If no category corresponding to the teaching defect image is found in the category list in step S63, a new category is additionally set in the category list in step S70. That is, as shown in FIG. 7, when the operator instructs the new category registration 48 (clicks with the mouse or inputs an execution command from the keyboard), the host computer 40 causes the new category to be input to the input column 49 by operating the keyboard 41. Enter information. 50 is instructed to complete the input, and 51 is instructed to interrupt or cancel the input. The completed new category is added to the category list 46. As a result, a new category is assigned to the input teaching defect image in step S71 and stored in the storage device 43. That is, the newly added category is registered in the storage device 43 as a category corresponding to the teaching defect image indicated by the teaching defect number.
[0035]
The defect image detailed information column 52 includes the defect image collection date and time, the process, the incidental status such as the wafer name, the total number of teaching input defect images to be assigned categories, the number of images that have been subjected to category assignment processing so far, the teaching defects Information incidental to the category input operation such as the frequency of the category assigned to the image is displayed. By instructing the teaching data display 53, the teaching data list shown in FIG. 8 is displayed for the category input completed so far. The category input is confirmed by the registration button 54. Further, by operating the scroll 55, the host computer 40 can call a plurality of defect images for teaching from the image recording device 34 back and forth in the order of defect numbers or in the order stored in the image recording device 34. Become. If a category has already been entered, that category can also be displayed. In addition, when a call is made for a defect image that is different from the defect image that is to be entered by inputting the category by an erroneous operation of the scroll 55, the category is not determined for the defect image that is being entered in the category, and the category is not yet determined. It remains registered. Further, when moving to another defect image and calling it from the image recording apparatus 34, in addition to using the scroll 55, the registered teaching data display 53 is executed, and the teaching data list shown in FIG. 8 is displayed. A method is also conceivable in which a column corresponding to an arbitrary defect number is designated.
[0036]
Next, step S65 and subsequent steps shown in FIG. 6 will be described. First, according to the first embodiment of the feature amount, if the registration of the input teaching defect image is confirmed in step S62, the host computer 40 issues a command to the image processing device 35 and the image recording device. 34. The defect image for teaching is read from 34, various m feature quantities are extracted by the image processing device 35, the various m feature quantities thus extracted are provided to the host computer 40, and the category assigned in step S64. Are stored in the storage device 43 as teaching data. Then, the host computer uses Ch in the feature vector space based on the set (c1, c2,..., Cm) of feature values of defect number h extracted (calculated) from the image processing device 35 as the teaching point coordinates. It can be calculated and confirmed. As shown in FIG. 10, for example, the teaching point coordinates Ch in the feature quantity vector space based on the axes c1 and c2 of the two feature quantities have the relationship expressed by the following equation (3).
Ch = (c1_h, C2_h(Equation 3)
Therefore, in step S65, the host computer 40 determines the closest input teaching point Cg (the distance between the two points is closest) based on the following equation (4) with respect to the input teaching point Ch. Whether or not the Euclidean distance between the input teaching point Ch and the searched existing teaching point Cg satisfies the condition that the threshold value THd or less is satisfied, based on the following equation (5). If this condition is not satisfied, there is no existing input teaching point Cg in the vicinity of the input teaching point Ch (the input check is not active = NO). The category assigned to the point Ch based on the category list is registered as it is. If the above condition is satisfied, the existing teaching point Cg exists near the inputted teaching point Ch (the input check is active = YES), and the process proceeds to step S66. By the way, the host computer 40 displays the searched defect image of the nearest input teaching point Cg in the display area 57 shown in FIG. 9, and the Euclidean distance between the teaching point Ch and the already input teaching point Cg, Whether or not the condition below the threshold value THd is satisfied is displayed in the column 58.
[0037]
[Expression 4]

[0038]
| CkCh | represents the Euclidean distance between the input teaching point coordinate Ch and the already input teaching point coordinate Ck. Cg is an already input teaching point closest to the input teaching point coordinate Ch. In the case shown in FIG. 10, the already input teaching point Cg closest to the input teaching point coordinate Ch becomes C1, and the Euclidean distance satisfies the threshold value.
[0039]
[Equation 5]

[0040]
As described above, when the host computer 40 determines in step S65 that the already input teaching point Cg exists in the vicinity of the input teaching point Ch (input check is active = YES), in step S66. The host computer 40 displays the teaching defect image of the nearest defect number g and the input teaching defect image of the defect number h together with the category searched for on the monitor 38 as shown in FIG. It should be noted that the category is assigned to the teaching defect image having the input defect number h in step S64. The teaching defect images displayed in this way may be displayed not only in the nearest teaching defect image but also in the number designated by the operator in ascending order of the distance from the input teaching point Ch. Next, in step S66, the operator assigns the category of the defect image for defect number h assigned in step S64 and displayed in the column 47 to the defect image for teaching of the nearest registered defect number g in the registered register. If the category is the same and the category is the same (if there is no other teaching point in another category in the vicinity), the teaching point Ch input in step S64 is assigned based on the category list. The category is registered as it is. If the categories are different as a result of the comparison, the teaching defect image having the defect number h is displayed in the display area 45 in step S67, and it is determined in step S68 that it is not necessary to register the teaching defect image for teaching. If so, the process returns to step S62 to cancel the registration by operating the registration cancel button 65. Accordingly, the process proceeds to step S69, and the next teaching defect image is called from the image recording device 34. If it is determined in step S68 that the teaching defect image needs to be registered for teaching, the category display column 47 is used to input a category that matches the category of the teaching defect image having the nearest defect number g. Will be registered.
[0041]
Next, in step S72, the process is repeated until the category is assigned to all the teaching defect images.
That is, using a plurality of feature amounts set in advance or a plurality of feature amounts extracted from a plurality of teaching defect image sets as a reference feature amount, using a multidimensional feature amount vector space formed by this reference feature amount, A defect classification attribute in which a defect image for teaching is associated with one point in the feature vector space by a reference feature amount, and a plurality of defect images for teaching are similarly taught in the feature vector space and the corresponding defect image data belongs (Category) is associated and stored. As described above, when the host computer 40 associates the teaching point Ch with the feature vector space and registers the teaching point Ch in the storage device 43 as teaching data, the operator uses the screen of the monitor 38 shown in FIG. Based on the inputted teaching defect image displayed in the display area 45 in steps S65 and S66, the neighborhood teaching defect image displayed in the display area 57 and the column 58 and its category are examined, and input in step S68. In the case where the registration of the defective image for teaching is to be performed with the already input category, it is determined whether or not the registration is continued in step S62, and in the case where the registration is performed by changing the category in step S68, Select from the category list as in step S64, or create a new category as in steps S70 and S71. Perform additional processing once again to register.
[0042]
As described above, the host computer 40 acquires the teaching data in a form in which the categories are properly divided in the feature vector space by executing the processes of steps S63 to S68, S70, and S71 with the monitor 38. Is possible. For example, as shown in FIG. 5B, in two feature quantity vector spaces, it is possible to acquire teaching data in a form in which categories are properly divided with × (category 1) and □ (category 2). It becomes.
By the way, the category set and registered is connected to the host computer 40 or information storage means (storage device) 43 connected to the host computer or the network 42 together with information that becomes an index of the defect image such as a defect number. Is recorded as teaching data in the above-described host system. The data content of the teaching data is as shown in FIG. The number m of feature amounts is a number determined by the number of feature amounts set in advance in the first embodiment of the feature amounts, which will be specifically described later, or the dimension number s of the eigenspace in the second embodiment of feature amounts. .
Furthermore, the contents displayed on the monitor 38 are shown in FIG. 45 is a display area for displaying an input defect image, and 47 is a column for displaying an input category for the input defect image. 57 is a display area for displaying near-input defect images, 58 is a category of near-input defect images, Euclidean distance between near-input defect images and input defect images, and the distance is equal to or less than a threshold value (THd). A column for displaying a status flag indicating whether or not the condition is satisfied, 59 is a button for scrolling through the defect images in order to view neighboring input defect images displayed in order of distance, 60 is a re-search execution instruction button, and 61 is a search Defect image count input column 62 is a category list display button or category list display itself, 63 is a new registration button or buttons and input display columns necessary for new category input indicated by 48, 49, 50 and 51 in FIG. The category display column of the input defect image shown in 47 accepts input, and accepts direct input to the column or input from 62 and 63. Reference numeral 64 denotes a button for executing category registration for an input defect image, and reference numeral 65 denotes a registration cancel button.
[0043]
Next, the classification process will be described. In step S36, the image processing device 35 similarly performs the above-described set of m feature amounts (c1, c2,...) For the defect image input and recorded as a classification target in the image recording device 34. The feature quantity vector is extracted as a feature quantity vector based on (cm) (hereinafter referred to as classification data), and the extracted feature quantity vector is provided to the host computer 40 and stored in the storage device 43. The stored classification data has the same contents as the data shown in FIG. 4C, but the category column 36 is blank. For this reason, for each defect for classification, the category is estimated from the feature amount using the classification parameter (teaching data) 27 described above. The estimated category is stored in the storage device 43 connected to the host computer 40. The stored (recorded) classification data 30 is transferred via the network 42 to the quality management system 7 which is a host system.
Next, the classification calculation (classification data creation process) performed in step S37 will be specifically described with reference to FIGS. 11, 12, and 13. FIG. First, the host computer 40 can form a feature vector space of each defect image to be classified by calling the feature vector of each input defect image to be classified from the storage device 43. Hereinafter, a point in the feature vector space represented by the feature vector generated from the defect image to be classified will be referred to as a classification target point C.
In step S102, the host computer 40 searches for the teaching point Ce that is closest to the classification target point C in step S102 for the defect number that is the classification target input in step S101, and the teaching corresponding to the search teaching point Ce. This is executed by setting the defect category recorded in the defect number e for use as a classification result.
[0044]
The Ci in the feature vector space of the defect number i calculated for the teaching defect image collected in the learning process is used as the teaching point coordinates, and is calculated in step S36 from the defect image to be classified collected in the classification process. The feature quantity vector, that is, C in the feature quantity space is set as a classification target point coordinate. In addition, the category assigned to the defect number i in step S32 for the teaching defect image collected in the learning process is Cat (i), and the classification defect image collected in the classification process is step S37. If the category that is classified and calculated by the above is described as Cat, the above classification process can be described by the following equation (6).
It is assumed that the teaching data is composed of n teaching defect images.
[0045]
[Formula 6]

[0046]
| CkC | represents the Euclidean distance between the teaching point coordinate Ck and the classification target point coordinate C. Specifically, it is represented by the following (Equation 7).
[0047]
[Expression 7]

[0048]
From the equation (6), the category of an arbitrary classification target can be uniquely determined in any case. That is, from the relationship of Equation (6), the category Cat of the classification target defect image having the classification target point coordinate C is the teaching defect in which the Euclidean distance | CkC | between the teaching point coordinate Ck and the classification target point coordinate C is minimum. It can be uniquely determined as the category Cat (i) assigned to the number i.
[0049]
However, even if the classification target point C is the nearest point, what is far from the registered teaching point Ci is not registered in the existing category, that is, it may be an unknown defect to be classified into a new defect category. There is also sex. Therefore, in step S102, the host computer 40 makes it impossible to classify unknown defects into existing categories, and it is appropriate to confirm the category to the operator. In order to identify such an unknown defect in the host computer 40, a threshold value THd for the Euclidean distance between the teaching point coordinates Ci and the classification target point coordinates C is set for the host computer 40 using the input means 43 or the like. The Next, the host computer 40 compares the set threshold value THd with the minimum value min | CkC | calculated in the formula (3) as shown in the following formula (8). If the category can be determined, the category is registered in the classification result table stored in the storage device 43 in step S103. If the category cannot be determined, the process proceeds to step S104. The defect number i to be classified is added to the classified unknown defect list stored in the storage device 43.
The classification unknown defect is a category in which the defect image i to be classified exists greatly away from the taught defect image in the feature vector space from the current teaching data stored in the storage device 43. Indicates something difficult to do.
[0050]
[Equation 8]

[0051]
As a result, when there is no teaching point Ci within (in the vicinity of) the threshold THd from the classification target point C in the feature vector space, the defect image of the classification target at the classification target point C has an unknown attribute, that is, an unknown defect. It becomes possible to classify as.
[0052]
Next, in step S105, the host computer 40 returns to step S106 until the classification category B is assigned to all the defect images for classification, and reads and repeats the next defect image for classification from the image recording device 34. become.
Further, in step S107, the host computer 40 checks whether or not the defect number is registered in the classified unknown defect list. If it is not registered, the host computer 40 ends. If it is registered, the host computer 40 classifies it as an unknown defect in step S108. The input defect image classified as an unknown defect (defect image of unknown category) is displayed on the monitor 38 shown in FIG. In step S109, the operator inputs and assigns an unknown defect category as necessary. When the category is presented by the operator, the category presented for this unknown defect is added to the feature vector of the unknown defect and recorded as teaching data.
In the recording process to the teaching data, in step S110, a new defect number is added to the teaching data list shown in FIG. 4C, and a category, a feature vector, and the like are added to the added defect number. The recording ends.
[0053]
That is, the category assignment to the classification unknown defect in step S110 will be described in detail below along the monitor display shown in FIGS. First, selection of a category unknown defect image to be a category assignment target is executed by a user selecting and instructing from the classified unknown defect images displayed in a list shown in FIG. Alternatively, input of a defect image file name using the keyboard 41 or selection from a defect image file name list may be used. On the monitor 38, the classification unknown defect image is displayed in the display area 66, and the teaching defect image at the neighboring point in the feature vector space is displayed in the display area 67 and the column 75 in order of distance together with the category. A similar image is selected from the defect images for teaching nearby points, and the category of the unknown defect is input to the column 73. The button 76 is used to scroll the neighboring point teaching defect images displayed in order of distance.
[0054]
By the way, when there is no defect image similar to the neighboring point teaching defect image, the category list displayed in the display area 76 is referred to, the corresponding category is selected, or a new category is set, and the classification unknown defect is determined. Give and register a new category. Selection from the category list or new category registration is the same as the method described in the learning process. That is, when the operator instructs the new category registration 77, the host computer 40 executes input of new category information to the input column 80 by operating the keyboard 41. 78 is indicated for completion of input, and 79 is indicated for interruption and cancellation of input. Further, the defect image detailed information column 74 is the same as that 52 described in the learning process.
Finally, the category input for the classification unknown defect image is confirmed by the button 69 for classification registration only. On the other hand, the registration button 68 confirms the category input for the classified unknown defect image, sets the classified unknown defect image as a teaching image, and newly sets the image and the feature amount obtained therefrom as learning data (teaching data). ) To add. Further, the categorization abandonment for the unknown classification defect in step S110 is executed by the delete button 70. Also, a plurality of classified unknown defect images are called back and forth by the button 71 in the list recording order. If a category has already been entered, that category is also displayed. When another classification / unknown defect image is called due to an erroneous operation of the button 71, the category is not determined for the classification unknown defect image during the category input operation, and the category classification is not performed. The function of the input check button 72 is as described in the learning process.
[0055]
By the way, the newly added defect number, corresponding category, and feature vector are collectively referred to as new teaching data, and a point in the feature vector space represented by the feature vector of the new teaching data is a new teaching point. I will call it. When the classification target point C is near the teaching point Ci as described above, the defect classification is performed by assigning the category of the teaching point to the classification target defect image. Therefore, even if a new teaching point is added, the position of the existing teaching point in the feature vector space does not change due to the addition of the new teaching point, so that the classification performance before the addition of the new teaching point is maintained. Is possible.
An exception is the case where a new teaching point with a different category falls within a distance of 2 × THd of the existing teaching point. Assume that an existing teaching point belongs to category P and a new teaching point belongs to category Q. In this case, before the new teaching data is input, all the classification target points C corresponding to the vicinity of the existing teacher points within the distance of THd are classified as the category P, but after the new teaching data is input, the nearest neighbor to be classified Depending on whether the teaching point becomes an existing teaching point or a new teaching point, a space region in which the category is P or Q can be formed in the feature vector space. As described above, such a case occurs when a new teaching point is placed within a distance twice the THd of an existing teaching point. However, the fact that the teaching data points having different categories are in the vicinity is inconsistent with the premise of the feature vector space designed to separate defects of different categories. When a new teaching point with a different category is placed within a distance of 2 × THd, the input defect image corresponding to the new teaching data and the input defect image corresponding to the existing teaching data are juxtaposed on the monitor 38 shown in FIG. And inquires of the operator whether there is any contradiction in teaching data input. The operator confirms the images juxtaposed on the monitor 38 and their categories, and selects either one of the categories, registration as a different category, registration or cancellation of registration as it is. As a result, it is possible to prevent inconsistent data from being input without being detected by the operator. Here, the threshold for determining whether or not to start such processing is 2 × THd, but α × THd is set so that a margin can be seen, and α can be set to an arbitrary value by the operator. It is also possible to allow external input.
[0056]
The so-called contradiction of registered teaching data described here is not a problem that occurs only in the learning process that is performed in parallel with the classification process, but may also occur in the initial learning process. For example, an input error of an operator (operator), teaching data input lacking in consistency of categories by an unskilled operator, and the like can be cited. However, such an erroneous input can be prevented by performing the same check as described above in the initial learning process.
The teaching data can be added while performing classification processing as described above, so the defect classification device follows the change in the correspondence between defect classification attributes (categories) and defect images caused by fluctuations in the semiconductor wafer manufacturing process. It becomes possible to make it. Further, it is possible to provide an image classification method and apparatus for semiconductor defects that can notify a user of a new defect occurrence by discriminating an unknown new defect caused by a process variation from a previously learned defect attribute (category).
The description has been made above assuming that the color information, shape, size, etc. of the defect are predetermined as the first embodiment of the feature amount. According to the first embodiment of these feature quantities, since they are color information, shape, size, etc., the meaning of the values is explicit for the operator, and the teaching data is appropriately stored in the feature quantity vector space. There is an advantage that it can be intuitively grasped whether the feature quantity extracted from the defect to be classified is appropriate. In particular, in the case of an optical image, it is possible to identify a category so that it can be assigned by extracting feature values such as color information, shape, and size.
[0057]
On the other hand, in the case of an image by secondary electrons or reflected electrons obtained by irradiating an electron beam, no color information can be obtained. Instead, the state of texture (subtle change in gray value), complex shape, etc. are explicitly indicated. It is necessary to identify so that a category can be assigned with a non-feature amount. However, in the case of a feature amount such as a texture state or a complicated shape, it is difficult to explicitly set, and therefore there is no guarantee that a category to be handled can be sufficiently identified. That is, even if a non-explicit teaching defect image such as a texture state or a complicated shape is displayed in the display area 45 of the monitor 38, the operator cannot easily determine which category A should be assigned.
[0058]
Therefore, a second embodiment of feature quantities that can solve this problem will be described.
Whether it is an optical image or an electron beam image, the teaching image signal of defect number k is represented by pixel signals (i1, i2,..., Ip) indicated by grayscale values obtained in the scanning order by raster scanning. It will be.
By the way, in order to calculate a feature quantity that is not affected by variations in sensitivity due to imaging conditions or the like from such teaching image signals, first, a normalized teaching image signal from which sensitivity variations due to imaging conditions or the like have been removed is obtained. Need to get. Accordingly, the image processing device 35 first determines the two-dimensional size (p = u × w) and brightness (gray value) for the teaching image signal of defect number k collected by the image recording device 34. Perform normalization. As a result, as shown in FIG. 14, a two-dimensional u × w that has been normalized to a certain two-dimensional size (p = u × w) in advance and further normalized for brightness (lightness value). A teaching image signal Ik (denoted by the following (Equation 9)) having a defect number k represented by a vector composed of gray values of pixels is obtained.
[0059]
[Equation 9]

[0060]
t is a symbol representing transposition of a matrix. p is the number of pixels.
As described above, the teaching image signal Ik is normalized to a predetermined size (p = u × w) with respect to the transposed matrix of [i1, i2,... (Value) is normalized as shown in the right term (sum of squares of gray value) in the above equation (9), and thus the normalized teaching that eliminates variations in sensitivity due to imaging conditions and the like An image signal for use can be obtained.
Therefore, in the second embodiment of the feature quantity, the teaching defect image is represented by an eigenspace (eigenvector space) spanned by eigenvectors of the covariance matrix. In order to calculate the eigenspace spanned by the eigenvector of the covariance matrix for the teaching defect image, first, the image processing device 35 first calculates an average image of all normalized teaching defect images collected by the image recording device 34. A is obtained by the following equation (Equation 10).
[0061]
[Expression 10]

[0062]
Next, the image processing device 35 subtracts the average image A from each of the teaching defect images (I1, I2,..., In), and generates a matrix X based on (Expression 11) shown below.
X≡ [I1-A, I2-A,..., In-A] (Equation 11)
Then, the image processing device 35 calculates the covariance matrix Q of the teaching defect image set based on the following equation (Equation 12).
Q≡X · (Transposed matrix of X) (Equation 12)
The eigenspace (subspace) spanned by the p-dimensional eigenvectors [e1, e2,..., es] is obtained by solving the following eigen equation (Equation 13).
[0063]
λq · eq = Q · eq (Equation 13)
That is, the feature quantity of the teaching defect image is an eigenspace (subspace) spanned by eigenvectors [e1, e2,..., Es] corresponding to s large eigenvalues (λ1 ≧ ... ≧ λs ≧ ... ≧ λp). Can be expressed. Note that s is a value smaller than p (= u × w).
As long as the value of s is increased, the feature quantity of the details of the teaching defect image can be expressed by the eigenvector [e1, e2,..., es], and if the value of s is small, the eigenvector [e1, e2,. Therefore, only the feature quantity of the rough image for teaching can be expressed. The feature quantity of the teaching defect image can be expressed by the eigenspace spanned by the eigenvector [e1, e2,..., Es].
The residual E between the original teaching defect image and the image expressed in the eigenspace of dimension s is expressed by the following equation (14), and the residual rate Re is expressed by the following (15). It is expressed by a formula.
[0064]
[Expression 14]

[0065]
[Expression 15]

[0066]
Therefore, the host computer 40 takes into account the classification performance such as the classification accuracy rate by using the residual rate Re calculated by the eigenvalue (λ1 ≧... ≧ λs ≧... ≧ λp) provided to the image processing device 35 as an index. S is determined and provided to the image processing apparatus 35.
As described above, the feature quantity in the second embodiment is a p-dimensional eigenvector [e1, e2,..., Es]. That is, the image processing device 35 calculates the eigenvector [e1, e2,..., Es] and provides it to the host computer 40. The calculation of the eigenvector may also be executed by the host computer 40.
[0067]
The host computer 40 uses the eigenvector provided from the image processing apparatus 35 or obtained by itself for the teaching defect image of the defect number k, the teaching point in the eigenvector space based on the following equation (16) It can be plotted at the position of i. FIG. 15 schematically shows the position of the teaching point i in one unique vector [e1] space. At this time, the host computer 40 needs to receive data relating to (Ik-A) from the image processing device 35 or the image recording device 34.
[0068]
ik = ([transposed matrix of [e1, e2,..., es]) × (Ik−A) (Equation 16)
Then, the host computer 40 stores the s elements of the vector ik and the category of the defect number k in the storage device 43 as the contents of the teaching data shown in FIG. In addition, eigenvectors [e 1, e 2,..., Es] (which are coordinate axes of feature amounts in the second embodiment) are also stored in the storage device 43 as feature amounts.
The above description is the feature amount extraction step S33 shown in FIG. 3A based on the second embodiment of the feature amount. Note that the teaching data creation step and the learning step S34 are executed in the same manner as in the first embodiment of the feature quantity, that is, as shown in FIGS. That is, the feature vector is [c, c,..., Em] in the case of the first embodiment of the feature amount, whereas the eigenvector [e1, e2 is in the case of the second embodiment of the feature amount. , ..., es].
Next, a classification process using the second embodiment of the feature amount will be described. The classification target defect image input from the image recording device 34 is normalized to the same size as the teaching defect image in the image treatment device 35, and the same normalization is applied to the brightness (lightness value). Let J be the classification target defect image vector after the normalization process. Next, in the image processing device 35 or the host computer 40, this vector J is projected onto a point j on the p-dimensional eigenvector space by the following equation (Equation 17).
[0069]
j = (transposed matrix of [e1, e2,..., ek]) × (JA) (Equation 17)
The category determination in step S102 shown in FIG. 11 is executed according to which point of the teaching point ik in which the classification target point j is scattered in the unique vector space. Searching for teaching points in the vicinity of the classification target point and category determination based thereon are processed based on the following equation (18) in the same manner as equation (6).
[0070]
[Formula 18]

[0071]
| Ikj | represents the Euclidean distance between the teaching point coordinate ik and the classification target point coordinate j. The processing of the classification unknown defect is the same as described above. In the learning process executed in parallel with the classification process, it is better to recalculate the eigenvectors [e1, e2,. Even if the eigenvector is recalculated, the teaching points are scattered in the s-dimensional space, and the category corresponding to the teaching points is unchanged, so that the continuity is guaranteed in the learning result.
[0072]
As described above, it is possible to accurately assign the classification category B to the defect image to be classified even if the second embodiment of the feature amount is used. In particular, in optical images such as excimer laser light having a short wavelength, or particle beam images obtained by irradiating an electron beam, color information cannot be obtained as feature amounts, and texture states (grayscale) Since a subtle change in value) and a complicated shape can be obtained, it is preferable to use the eigenvector space according to the second embodiment as the feature vector space.
[0073]
【The invention's effect】
According to the present invention, since the data for teaching can be added while performing the defect image classification processing, the change in the correspondence between the defect classification attribute and the defect image caused by the variation of the semiconductor wafer manufacturing process can be reduced. There exists an effect which can make classification follow.
Further, according to the present invention, an unknown new defect caused by a process variation can be separated from the already learned defect attributes, and a new defect occurrence can be notified to the user. Since it is possible to reliably detect a new defect occurring in the process, there is an effect that it is possible to realize defect information collection for highly reliable process management.
[Brief description of the drawings]
FIG. 1 is a diagram for explaining a role of a defect classification method and its apparatus according to the present invention in a semiconductor manufacturing process.
FIG. 2 is a block diagram showing an embodiment of a defect classification apparatus according to the present invention.
FIG. 3 is a diagram for explaining respective processing procedures of a learning process and a classification process performed by the defect classification apparatus according to the present invention.
FIG. 4 is a diagram showing a teaching data creation procedure, a monitor screen, and teaching data in a learning process according to the present invention.
FIG. 5 is a diagram for explaining the principle of learning according to the present invention.
FIG. 6 is a flowchart specifically showing a procedure for creating teaching data according to the present invention.
7 is a diagram showing a monitor screen for executing the teaching data creation procedure shown in FIG. 6; FIG.
FIG. 8 is a diagram showing an example of a category list that is teaching data according to the present invention.
FIG. 9 is a diagram showing a monitor screen for executing the teaching data creation procedure shown in FIG. 6;
FIG. 10 schematically shows the relationship between the input teaching point coordinates and the already input teaching point coordinates in the first embodiment of the feature amount vector space based on two feature amounts in the teaching defect image according to the present invention. It is a figure.
FIG. 11 is a flowchart specifically showing a classification processing procedure according to the present invention.
FIG. 12 is a diagram in which a teaching defect image of a teaching point existing in the vicinity of a classification target point in the feature vector space is output and displayed on a monitor screen when a classification category is assigned to the classification target defect image. is there.
FIG. 13 is a diagram showing a monitor screen for executing a classification processing procedure according to the present invention.
FIG. 14 is a diagram showing a gray value teaching defect image obtained by raster scanning according to the present invention.
FIG. 15 is a diagram showing teaching points in the eigenvector space in the second embodiment of the feature vector space according to the present invention.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... Probe inspection process, 2 ... Appearance inspection apparatus, 3 ... Defect classification apparatus, 4 ... Appearance inspection result, 5 ... Classification object defect image, 6 ... Screen which shows generation frequency according to defect mode, 7 ... Quality control system, 8a 8b ... Substrate, 27 ... Classification parameter (teaching data), 28 ... Substrate transport device, 29 ... Optical system, 30 ... Classification data, 31 ... Stage, 32 ... Imaging device, 33 ... Image input device, 34 ... Image recording Device (image storage device), 35 ... Image processing device, 36 ... Stage control unit, 37 ... Substrate transport control unit, 38 ... Monitor, 40 ... Host computer, 41 ... Input means, 42 ... Network, 43 ... Storage device, 45 ... Display area for displaying input defect image, 47 ... Column for displaying input category, 52 ... Detailed information column for defect image, 57 ... Display area for displaying near-input defect image, 58 ... Column 59 ... button, 62 ... category list display button, 66 ... display area such as classified unknown defect image, 67 ... display area such as teaching defect image, 73 ... category input column, 74 ... defect image detailed information column, 75 ... category Display column.

Claims (5)

Teaching data indicating the correspondence between a plurality of defect images for teaching and categories corresponding to the types of defects for classifying the types of defects based on the classification target defect images obtained by imaging the inspection object Is a method of creating
A plurality of types of image feature amounts are calculated in advance for each of a plurality of already input teaching defect images, and a correspondence relationship between the plurality of already input teaching defect images and a category corresponding to the type of defect is shown. A first step of creating teaching data;
A plurality of types of image feature amounts are calculated for a newly input teaching defect image h, and the new input is calculated for the teaching data created in the first step. A second step of assigning a category corresponding to the type of defect based on the defect image for teaching h;
The new input teaching point coordinates Ch in the multi-dimensional image feature vector space based on a plurality of types of image feature values for the new input teaching defect image h calculated in the second step and the first step. A plurality of distances between the plurality of already input teaching point coordinates in the multi-dimensional image feature vector space based on a plurality of types of image feature values for each of the plurality of already input teaching defect images calculated in Among them, the already input teaching point coordinate Cg of the nearest neighbor distance is searched for the newly input teaching point coordinate Ch, and if the searched nearest neighbor distance is larger than a predetermined threshold value THd, a new one is found. If the category assigned in the second step is directly adopted as the category for the input teaching defect image h, and the searched nearest neighbor distance is less than or equal to a predetermined threshold value THd, The category assigned to the newly input teaching defect image h in the second step is compared with the category assigned to the already input nearest-neighbor distance teaching defect image obtained from the teaching data. When the comparison results are the same, it is necessary to adopt the same category as the category for the newly input teaching defect image h, and to register the newly input teaching defect image h when the comparison results in a difference. In some cases, there is provided a third step of adopting a category assigned to the previously input nearest-neighbor distance teaching defect image as a category for the newly input teaching defect image h. Data creation method . 2. The teaching data generation method according to claim 1, wherein the image feature vector space is an eigenvector space obtained by performing eigenvalue conversion on a defect image feature . A plurality of types of image feature amounts are calculated in advance for each of a plurality of already input teaching defect images, and a correspondence relationship between the plurality of already input teaching defect images and a category corresponding to the type of defect is shown. A first step of creating teaching data;
For a newly input teaching defect image h, a plurality of types of image feature amounts are calculated, and the previously input teaching data created in the first step is A second step of assigning a category corresponding to the type of defect based on the newly input teaching defect image h;
The new input teaching point coordinates Ch in the multi-dimensional image feature vector space based on a plurality of types of image feature values for the new input teaching defect image h calculated in the second step and the first step. A plurality of distances between the plurality of already input teaching point coordinates in the multi-dimensional image feature vector space based on a plurality of types of image feature values for each of the plurality of already input teaching defect images calculated in Among them, the already input teaching point coordinate Cg of the nearest neighbor distance is searched for the newly input teaching point coordinate Ch, and if the searched nearest neighbor distance is larger than a predetermined threshold value THd, a new one is found. If the category assigned in the second step is directly adopted as the category for the input teaching defect image h, and the searched nearest neighbor distance is less than or equal to a predetermined threshold value THd, The category assigned to the newly input teaching defect image h in the second step is compared with the category assigned to the already input nearest-neighbor distance teaching defect image obtained from the teaching data. When the comparison results are the same, it is necessary to adopt the same category as the category for the newly input teaching defect image h, and to register the newly input teaching defect image h when the comparison results in a difference. A third step of adopting the category assigned to the previously input nearest-neighbor distance teaching defect image as the category for the newly input teaching defect image h in some cases ;
A teaching data creation process for creating teaching data indicating a correspondence relationship between a plurality of teaching defect images having a category corresponding to a defect type;
The classification target defect image obtained by imaging the object to be inspected can be classified into a known classification category based on the teaching data created in the teaching data creation process, and cannot be classified. In some cases, the classification process is stored as information about unidentified defects that cannot be classified.
A defect classification method characterized by comprising: A plurality of types of image feature amounts are calculated in advance for each of a plurality of already input teaching defect images, and a correspondence relationship between the plurality of already input teaching defect images and a category corresponding to the type of defect is shown. A first step of creating teaching data;
For a newly input teaching defect image h, a plurality of types of image feature amounts are calculated, and the previously input teaching data created in the first step is A second step of assigning a category corresponding to the type of defect based on the newly input teaching defect image h;
The new input teaching point coordinates Ch in the multi-dimensional image feature vector space based on a plurality of types of image feature values for the new input teaching defect image h calculated in the second step and the first step. A plurality of distances between the plurality of already input teaching point coordinates in the multi-dimensional image feature vector space based on a plurality of types of image feature values for each of the plurality of already input teaching defect images calculated in Among them, the already input teaching point coordinate Cg of the nearest neighbor distance is searched for the newly input teaching point coordinate Ch, and if the searched nearest neighbor distance is larger than a predetermined threshold value THd, a new one is found. If the category assigned in the second step is directly adopted as the category for the input teaching defect image h, and the searched nearest neighbor distance is less than or equal to a predetermined threshold value THd, The category assigned to the newly input teaching defect image h in the second step is compared with the category assigned to the already input nearest-neighbor distance teaching defect image obtained from the teaching data. When the comparison results are the same, it is necessary to adopt the same category as the category for the newly input teaching defect image h, and to register the newly input teaching defect image h when the comparison results in a difference. A third step of adopting the category assigned to the previously input nearest-neighbor distance teaching defect image as the category for the newly input teaching defect image h in some cases;
A teaching data creation process for creating teaching data indicating a correspondence relationship between a plurality of teaching defect images having a category corresponding to a defect type;
A plurality of types of defect image feature amounts are calculated for the classification target defect image obtained by imaging the inspection target object, and a multidimensional type based on the plurality of types of defect image feature amounts of the calculated classification target defect image is calculated. A plurality of teaching defect images in the multi-dimensional image feature vector space created in the teaching data creation process and registered as teaching data for the classification target point coordinates for the classification target defect image in the image feature vector space Among the plurality of distances between the plurality of teaching point coordinates, the teaching point coordinates of the nearest neighbor distance with respect to the classification target point coordinates are searched, and the searched nearest neighbor distance is a predetermined threshold. A classification value is determined to determine whether or not the classification target defect image can be classified into a known classification category obtained from the teaching data. And the classification process of storing as information
A defect classification method characterized by comprising: 5. The defect classification method according to claim 3, wherein the image feature vector space is an eigenvector space obtained by performing eigenvalue conversion on a defect image feature .

Images