JP4024381B2 - Defect inspection method and apparatus - Google Patents

Description

【００４９】
位置ずれ検出は、上記のＳ（Δｘ、Δｙ）をｍｉｎとするΔｘ、Δｙを検出するものである。
【００５０】
ただし、最小となる位置は画素単位でしか求められないため、真の位置が求めたΔｘ、Δｙのどちらの近くにあるかにより、オフセットとして加える。
【００５１】
下記の式に基づき、Δｘ、Δｙに１を加えたりもしくは、そのままにする。
【００５２】
すなわち、
【００５３】
【数２】

【００５４】
【数３】

【００５５】
【数４】

【００５６】
【数５】

【００５７】
なお、Δｘ＋＋は、Δｘ＝Δｘ＋１の意である。
【００５８】
このように位置合わせにより、つねに得られた位置ずれ量だけ片方の画像の位置をずらせることにより、二枚の画像の位置合わせが行なわれる。すなわち、画像ｆを常に右上移動して、新たな画像ｆ′を求めることになり、移動方向を４種類(右下移動、左上移動、左下移動、右上移動）から１種類に特定することができる。これはハードウエアの簡単化につながるものである。
【００５９】
図３は、明るさ一致フィルタ操作部１２を詳しく示したものである。画素単位で位置合わせされた２枚の画像ｆ（ｘ、ｙ）とｇ（ｘ、ｙ）に関して、まず画像内で下記値を最小にするフィルタＦ、Ｆ‘を求める。
【００６０】
【数６】

【００６１】
フィルタＦ、Ｆ‘は、例えば２×２画素のサイズをもつものである。
【００６２】
図５にフィルタの１例を示す。フィルタＦとＦ‘は、同図に示すように対称形であり、双対である。このような双対形にすると、フィルタのパラメータの係数を、最小二乗法を用いて、解くことができる。
【００６３】
【数７】

【００６４】
【数８】

【００６５】
ただし、
【００６６】
【数９】

【００６７】
【数１０】

【００６８】
【数１１】

【００６９】
本方式は、対象である二枚の画像にフィルタリングを施し、その濃淡の二乗誤差を最小にして一致させるものである。繰返し演算などが不要であり、一回の演算で実現可能なものである。
【００７０】
上記方式の特徴は、二枚の画像の濃淡が、二乗の誤差最小の意味でよく一致するようにフィルタ係数α、βを求めるものである。特に、これらのパラメータは必ずしも画像の位置ずれ量の真値を表しているわけではない。例えば、従来技術において述べたように、Ｓ（Δｘ、Δｙ）に放物面を当てはめ、その最小位置を算出し、次にこの算出位置に基づいて、内挿して補間画素を求める方式も考えられるが、この場合は明るさに関して、なんら満たすべき規範とか条件はなく、その意味で、得られる画像を比較検査に使うことは、保証されたものでない。さらに、明るさが異なる場合は、算出した位置ずれは何を表しているか不明である。その上、たとえ放物面近似して算出した最小位置が、本実施例で述べた方式と位置が一致しても、得られる比較すべき画像は一致するものではない。
【００７１】
提案したマッチング方式は、２枚の画像の明るさの２乗誤差が最小になることを保証するものであり、この点で他の方式と一線をかくすものである。図１４に示すように、線形近似であるため、フィルタの係数であるαは、位置ずれとみなすと、誤差をもつ。しかし、得られた明るさは一致している。 本方式は、画像間の濃淡の違いを本質的に小さくでき、比較検査において極めて都合のよい方式であると考えられる。
【００７２】
さらに、フィルタ係数α、βの算出は、繰返し演算することなく解析的に行うことができ、ハードウエア化するのに適しているという特長がある。
図４は、局所的な階調変換部１３を詳しく示したものである。画素単位で位置合わせされ、明るさ一致フィルタ操作出力の２枚の画像ｆ（ｘ、ｙ）とｇ（ｘ、ｙ）に関して、まず画像内の定めたエリア内で下記の値を最小にするパラメータａ、ｂ（a：ゲイン、ｂ：オフセットに相当）を求める。
【００７３】
【数１２】

【００７４】
パラメータａ、ｂは、上記した値をａ、ｂで偏微分し、零とおくことによって求められる。例えば、定めたエリアとは、各点の周囲の７の範囲である。
【００７５】
求められたパラメータを用いて画像信号の片方ｇ（ｘ、ｙ）を
【００７６】
【数１３】

【００７７】
と変換する。これにより明るさが一致した画像が得られる。パラメータａ、ｂは、位置（ｘ、ｙ）ごとに異なる値を取りうる。
【００７８】
【数１４】

【００７９】
【数１５】

【００８０】
ここで、Σの範囲の画素数をＭＮとしている。
【００８１】
さらに、Σの範囲において、着目した中央画素と周囲画素の明るさを比較し、明るさが大きく異なるときは、これを加算しないこともよい方法である。
【００８２】
或いは、加算自体はするのであるが、加算するとき、重みを付けて、寄与率を下げることが有効である。例えば、（ｘ、ｙ）の着目画素の明るさをｃとし、Σの範囲内の別の画素の明るさをｄとすると、重みＷ（ｘ、ｙ）を
【００８３】
【数１６】

【００８４】
とする。ここで、ｍａｘ[ ]は最大値検出、ｃ、ｄはいずれも８ビット階調の明るさである。また、Ｄは定数とする。これにより、着目した中央画素と周囲画素の明るさが似ているときは、重みを１近くの値にし、似ていないときは、小さい値とするものである。Dは定数としたが、明るさに応じて可変してもよく、D＝func（ｃ）である。さらには、同じパターンに属すか否かで判断し、異なるパターンの平均明るさをμとすると、D＝｜ｃ−μ｜として与えてもよい。パターンの種類が３つ以上ある場合には、似ているパターン間の明るさの違いをＤとすればよい。勿論、このような形態にこだわる必要は必ずしもなく、重みを適切に与えられるものならばよい。
【００８５】
図６に検出した２枚の画像の例を示す。検出した２枚の画像には、図示したような明るさに大きな違いがある。この２枚の画像を、画素単位の精度で位置合わせを行い、明るさ一致フィルタ操作をした。しかしながら、本画像では明るさの違いが大きすぎて、同図に示すように、差画像において、大きな不一致が生じている。このような画像に対し、階調変換処理を実施した。
【００８６】
図７〜９に処理例を示す。図７〜９において、検出した２枚の画像ｆ（ｘ、ｙ）、ｇ（ｘ、ｙ）、及び明るさヒストグラムを示す。また、変換後の画像ａ＊ｇ（ｘ、ｙ）＋ｂ、及び明るさヒストグラムを示す。ここでは、Ｄ＝７０とした。図７に示すヒストグラムからわかるように、Dの値は双峰性のヒストグラム分布の各分布平均の差に相当する値になっている。即ち、このようなDが設定された重みWは、同じ分布に属するかどうかの指標になっている。なお、定めたエリアは、各点の周囲の７×７画素の範囲である。図７〜９から、変換によって明るさヒストグラムがほぼ等しくなることがわかる。ここで、図７〜９の画像を対象にした実験結果では、画像内のある点でａ＝１．４１、ｂ＝０が得られた。この画像では、明るさのゲインが大きく（４１％）異なることがわかる。
【００８７】
この例から、オフセットｂを常に０と設定し、ゲインだけを可変にする方式も考えられる。これらは、対象とするパターンの特性や装置構成等により、決定すればよい。
【００８８】
図１０及び図１１は、このような変換により得られた画像の差を示したものである。図１０及び図１１において、最初の３つは、定めたエリアを、各点の周囲の３×３、５×５、７×７画素の範囲としたものである。また、この時の重みは、Ｗ（ｘ、ｙ）＝１とした。さらに、最後の１つは、７×７画素の範囲について、重みを上記したＷ（ｘ、ｙ）に従うとしたものを示した。これらの図より、エリアが小さい場合は、明るさを局所的に合わせ込み、画像間の不一致が小さくなることがわかる。これでは、明るさの許容範囲は広がるが、微小な欠陥を見逃してしまう。このため、エリアを検出すべき欠陥に応じて拡げる必要がある。しかし、重みが１に固定されている場合、パターン領域の境界部が不一致として、即ち虚報として検出されてしまう。重み付けした場合は、境界の影響が低減され、ほぼ２枚の画像の明るさが一致し、微小な欠陥も検出できることになる。
【００８９】
なお、７×７画素などのエリアは、必ずしも正方形である必要性はなく、長方形や多角形、或いは円でもよい。また、このような極微小な領域でなく、数１００×数１００画素といった領域にしてもよく、要は明るさの変動が吸収可能な範囲であればよい。
【００９０】
重み付けの方法は、この他にも、着目した中央画素と周囲画素の明るさを比較し、明るさの差がしきい値より大きいときは、重みを０とすることもできる。
【００９１】
以上述べた以外に、下記の階調変換も考えられる。
【００９２】
【数１７】

【００９３】
ここで、σｆ、ｍｆは、それぞれ画像ｆ（ｘ、ｙ）において、点（ｘ、ｙ）近傍の定めたエリア内の標準偏差、平均値であり、σｇ 、 ｍｇは、それぞれ画像ｇ（ｘ、ｙ）において、点（ｘ、ｙ）近傍の定めたエリア内の標準偏差、平均値である。
上記した変換により、画像ｇ（ｘ、ｙ）の明るさをｆ（ｘ、ｙ）に一致させることができる。
【００９４】
Ｗ（ｘ、ｙ）は、上記した重みでもよいし、画像ｆ（ｘ、ｙ）とｇ（ｘ、ｙ）の定めたエリア内の画像データの相関係数でもよい。
【００９５】
この方式も、結果として２つの画像のヒストグラムが一致するという特長がある。
【００９６】
いずれの方式も、ゲインとオフセットからなる線形変換の形になっている。
【００９７】
以上のべた階調変換は、着目画素近傍の局所的な明るさ変換である。勿論、目的や画像の特性に従って、画像全体、即ちここでは２５６ライン全体に適用しても構わないものである。また、２枚の画像のどちらに、一方の画像の明るさを一致させるかは、例えば、画像毎に明るさ平均値を算出し、明るい方に合わせることができる。また、定めたエリア毎、或いは各点毎に、明るさ平均値を算出し、明るい方に合わせることもできる。
【００９８】
なお、図１では、画像明るさ一致フィルタ操作の後に、階調変換を実施しているが、この順序は、図２に示すように入れ替えてよい。
【００９９】
次に、比較器１４は、本発明者らが開発した方式、特開昭６１−２１２７０８号公報記載の技術に示したもの等で差し支えなく、画像の差画像検出回路、差画像を定めたしきい値で２値化する不一致検出回路、２値化された出力より面積や長さ（投影長）、座標などを算出する特徴抽出回路から構成される。
【０１００】
図１２及び図１３を用いて、さらに本発明にかかる２値化の際のしきい値設定について説明する。
【０１０１】
上記したように、差画像を２値化する際、領域の境界などで、虚報が発生しやすい。そこで、図１２に示すように、画像の局所領域内の最大値と最小値の差、平均値、微分値のｘ、ｙいずれか大きい方（以下、局所データ）を、各点ごとに算出する。これらの値に、別途定めたパラメータを掛け合わせ、加算する、いわゆる積和演算することにより、しきい値を生成する。これにより、例えば、領域の境界部の明るさ変化が大きいところでは、微分値等が大きくなるので、しきい値が大きくなり、虚報の発生を未然に防ぐことが可能になる。勿論、最大値と最小値の差、平均値、微分値のｘ、ｙいずれか大きい方は、３つとも備える必要はなく、いずれかひとつでもよい。例えば、上記した階調変換を実施した場合は、平均値を算出する必要性はない。
【０１０２】
このようなしきい値を用いて、画像の差を２値化すれば、虚報の問題をより効果的に低減できる。なお、上記局所データは、後述する散布図により分布を求めることにより、その大小の決定支援が可能である。図３１〜３３には、画像の局所領域内の最大値と最小値の差に関する散布図を示している。この分布データに線分を当てはめ、線分からの誤差を求める。このような処理を各局所データに関して行い、その積和によってしきい値を決めることが可能である。
【０１０３】
たとえば、しきい値Ｔｈとして
Ｔｈ＝Ｃ３×局所コントラスト＋Ｃ２×明るさ平均
を選ぶと考える。ここで、局所コントラスト画像を、各画素に関して３×３画素の最大値−最小値
で定義し、明るさ平均画像を、各画素に関して３×３画素の移動平均であらわす。
【０１０４】
比較する２枚の画像の局所コントラスト画像をｆ（ｘ、ｙ）、ｇ（ｘ、ｙ）とし、
【０１０５】
【数１８】

【０１０６】
上式より求められるＶｅをσｋとする。
【０１０７】
同様に、明るさの平均画像をｆ（ｘ、ｙ）、ｇ（ｘ、ｙ）とし、求められるＶｅをσａとする。
【０１０８】
これらより、下式の（数１９）が決まる。
【０１０９】
【数１９】

【０１１０】
同様のことを、画像を変えて行い、係数のＣ２、Ｃ３が求められる。
【０１１１】
上記したＴｈを選ぶと考えると、局所コントラストの散布図において、傾き１（ｍ＝１）、切片０（ｎ＝０）の直線から各プロットデータ点までの距離であり、誤差に相当する標準偏差σｋを求める。同様に、明るさ平均の散布図において、傾き１、切片０の直線から各プロットデータ点までの距離であり、誤差に相当する標準偏差σaを求める。また、もとの２枚の画像の明るさの散布図において、傾き１、切片０の直線から各プロットデータ点までの距離であり、誤差に相当する標準偏差σｇを求める。
【０１１２】
これらの値を上記Ｔｈの式に代入すると、（数１９）のような、Ｃ２とＣ３の式が得られる。この操作を異なる場所の画像について行うと、係数が異なる別のＣ２とＣ３に関する式が得られる。従って、これらを連立方程式として解くと、Ｃ２、Ｃ３という係数が一意に決まる。これより、C2、C3からしきい値Thが決められる。もちろん、しきい値Thは、
Ｔｈ＝Ｃ３×局所コントラスト＋Ｃ２×明るさ平均＋オフセット
としてもよい。
【０１１３】
更に、別の設定方式として、求めるべき浮動しきい値を、明るさの局所的なコントラストや平均値の線形結合として（数２０）で表し、比較する２枚の画像の散布図情報を用いて重回帰分析によりパラメータを算出してもよい。
【０１１４】
【数２０】

【０１１５】
この時の設定の手順を以下に示す。
【０１１６】
（１）複数点の画像（２チップ一組）を検出する。
【０１１７】
（２）検出画像、参照画像のデータから明るさの散布図を作成する。（欠陥を含まない画像を対象とするか、または、欠陥部を除いた画像を用いる。）
（３）散布図におけるデータの集合を包絡する点を求めて（評価では、頻度１の点を抽出する）、その点に対応する画像の画素から、局所的コントラストや平均値のデータを抽出する。
【０１１８】
（４）上記（３）で求めた情報をもとに重回帰分析によりパラメータＣ０〜Ｃ４を調整する。
【０１１９】
（５）ｐ値（有意水準）に応じて用いるデータの選択を行う（ｐ値が十分信頼の持てる値（０.０５以下）となる組み合わせを求める）。
【０１２０】
（６）求めたパラメータＣ０〜Ｃ４から、しきい値画像を算出して差画像と比較を行う。
【０１２１】
（７）虚報があれば虚報の情報の加えて、パラメータＣ０〜Ｃ４を調整する。
【０１２２】
（８）テスト検査を行う。
【０１２３】
（９）虚報が発生したら、上記（７）と（８）とを繰り返す。
【０１２４】
また、図１３に示すように、上記したように係数と誤差の積和演算ではなく、ルックアップテーブル（ＬＵＴ）を用いても構成可能である。図１２及び図１３に示すように、画像の局所最大値と局所最小値を求め、その差よりコントラストを求め、これをＬＵＴに入力する。同様に、画像の局所平均値を求め、ＬＵＴに入力する。これらのＬＵＴ出力を別のＬＵＴに入力し、しきい値を求める。ＬＵＴの規模を適切なものとするため、図１２及び図１３では使用するビット数を８→６ビットに制限している。求めたしきい値は、比較回路（図１及び図２）１４に入力し、２値化のしきい値として使用する。ＬＵＴの中身のデータは、種々の画像を用いて、上記した手順と同様の手順で、誤差等を求め、これを補間等してデータを埋めて作成する。
【０１２５】
選ぶ画像は、誤検出しやすい場所が適していることは言うまでもないが、従来はこのような決め方がなく、本発明により決め方の手順が決まっただけでなく、きわめて論理的に決められるということが特徴である。
【０１２６】
図１及び図２において、設計情報に基づいて得られる半導体ウエハ４上におけるチップ内の配列データ等の座標を、キーボード、ディスク等から構成された入力手段１５で入力しておくことにより、ＣＰＵ１６は、入力された半導体ウエハ４上におけるチップ内の配列データ等の座標に基づいて、欠陥検査データを作成して記憶装置１７に格納する。
【０１２７】
この欠陥検査データは、必要に応じてディスプレイ等の表示手段に表示することもできるし、また出力手段により出力することもできる。
【０１２８】
また、このデイスプレイ上に、階調変換する前の画像または画像データと階調変換後の画像または画像データとをそれぞれ表示することにより、または、階調変換後の画像または画像データを表示することにより、階調変換が正しくなされて検査が行われていることを、作業者が目視で確認することができる。
【０１２９】
これらにより、きわめて高精度に画像を比較することができ、本発明で意図した目的がより高感度に実現できることになる。
なお、本実施例では明視野照明も用いた例を示したが、暗視野照明により得られた画像を対象にしてもよい。また欠陥の種類もショートやオープンなどの形状不良や、それ以外の異物も対象となる。
【０１３０】
〔実施の形態 ２〕
本発明に係るパターン検査方法および装置の第２の実施の形態を図１９に示す。ここでは、電子線により試料を走査して、電子線の照射によってウエハから発生する電子を検知し、その強度変化に基づいて走査部位の電子線像を得、電子線像を用いてパターン検査を行う。前述の発明が解決しようとする課題に対して、本第２の実施の形態では、欠陥判定のしきい値を、パターンの位置ずれ、階調値の差異などを考慮して画素ごとに設定することで対処する。
【０１３１】
本システムは、検出部１０１、画像取り出し部１０２、画像処理部１０３、システム全体を制御する全体制御部１０４からなる。
【０１３２】
始めに、検出部１０１について述べる。
【０１３３】
図１９において、電子銃３１を出た電子ビームは、磁界レンズ３２、対物レンズ３３を経て、試料面では画素サイズ程度のビーム径に絞られる。この際、グランド電極３７、リターディング電極３８によって、試料に負電位を引加し、対物レンズと試料間で電子ビームを減速することで低加速電圧領域での高分解能化を図る。電子線が照射されると、試料（ウエハ１）からは電子が発生する。偏向器３４による電子線のＸ方向の繰り返し走査と、ステージ２による試料のＹ方向の連続的な移動に同期して試料から発生する電子を検出することで、試料の２次元の電子線像が得られる。試料から発生した電子は検出器３５で捕らえられ、アンプ３６で増幅される。ここで、高速検査を可能にするために、電子ビームをＸ方向に繰り返し走査させる偏向器３４としては、偏向速度の速い静電偏向器を用いることが、また、電子銃３１としては、電子ビーム電流を大きくできるので照射時間が短くできる熱電界放射型電子銃を用いることが、また、検出器３５には高速駆動が可能な半導体検出器を用いることが望ましい。
【０１３４】
次に、画像取り出し部１０２について述べる。
【０１３５】
アンプ３６で増幅された信号は、ＡD変換器３９によってディジタル信号に変換された後、前処理回路４０において、暗レベル補正（ビーム・ブランキング期間中の特定数画素の階調値の平均を暗レベルとする）、電子ビーム電流の揺らぎ補正（ビーム電流を図示されていない対物絞りで検出し、信号をビーム電流で正規化する）、シェーディング補正（ビーム走査位置による光量変動を補正）を行った後、ガウシアンフィルタ、平均値フィルター、あるいはエッジ強調フィルターなどによるフィルタリング処理を行って、画質を改善する。また、必要に応じて、画像の歪みの補正も行なう。これらの前処理は、後の欠陥判定処理において有利なように検出画像を変換するためのものである。
【０１３６】
遅延回路４１は信号を一定時間だけ遅延させるが、遅延時間を、例えばステージ２がチップピッチ分移動する時間にすれば、遅延された信号g0と遅延されていない信号f0は隣り合うチップの同じ箇所での画像信号となり、前述したチップ比較検査となる。あるいは遅延時間を、ステージ２がメモリセルのピッチ分移動する時間にすれば、遅延された信号g0と遅延されていない信号f0は隣り合うメモリセルの同じ箇所での画像信号となり、前述したセル比較検査となるわけである。以上のようにして、画像取り出し部１０２から、比較すべき画像信号f0とg0が取り出される。以下、f0を検出画像、 g0を比較画像と呼ぶことにする。
【０１３７】
次に、画像処理部１０３における処理について述べる。
【０１３８】
画素単位の位置合わせ部４２では、検出画像の位置を基準にした時、検出画像と比較画像の「整合度」が最大となる位置が、０〜１画素の間になるように、比較画像の位置をずらす。
【０１３９】
次に、明るさ一致フィルタ操作部における、画像間あかるさ不一致を最小にするフィルタＦ、Ｆ‘を求める。前述したように、フィルタのパラメータの係数を、最小二乗法を用いて、解いて(7)式、(8)式のように、 dx0、dy0を求めるには種々の統計量ΣΣ××を求める必要がある。統計量算出部４４において、それら種々の統計量の算出を行い、サブＣＰＵ４５において計算されたΣΣ××を用いて、(数7)、(数8)の演算を行ってα、βを求める。
【０１４０】
局所階調変換回路４６は、前述したf1とg1の明るさをそろえる階調変換を実施するためのものである。
【０１４１】
差分抽出回路４９では、f1とg1の差画像sub(x,y)を求める。式で表すと次のようになる。
【０１４２】
【数２１】

【０１４３】
しきい値演算回路４８では、遅延回路４６、４７を経た画像信号f1、g1および、α、βを用いて、sub(x,y)の値に応じて欠陥候補か否かを判定する二つのしきい値thH(x,y)とthL(x,y) を算出する。 thH(x,y)はsub(x,y)の上限を規定するしきい値であり、 thL(x,y)はsub(x,y)の下限を規定するしきい値である。しきい値演算回路４８の構成を図２０に示す。演算の内容を式で表すと次のようになる。
【０１４４】
【数２２】

【０１４５】
【数２３】

【０１４６】
ただし、
【０１４７】
【数２４】

【０１４８】
【数２５】

【０１４９】
【数２６】

【０１５０】
ここで、 a a、 b bは、0〜0.5の実数、γは０以上の実数、εを０以上の整数
【０１５１】
【数２７】

【０１５２】
しきい値を算出する（数２２）（数２３）の右辺の第１項A(x,y)は、位置ずれ検出部４３で求めたα、βに対応してしきい値を補正するための項である。例えば（数２７）で表されるdx1は、f1の階調値のｘ方向の局所的な変化率とみなし、 dx1(x,y) *αは、位置がαずれた時のf1の階調値の変化の予測値ということができる。よって、A(x,y)の第１項{dx1(x,y) *α - dx2(x,y)*(- α)}は、ｘ方向にf1の位置をα、g1の位置を- αずらした時に、f1とg1の差画像の階調値がどのぐらい変化するかを画素ごとに予測した値ということができる。同様に第２項はｙ方向について予測した値ということができる。 しきい値の第一項A(x,y)は、α、βをキャンセルするための項である。
【０１５３】
しきい値を算出する（数２２）（数２３）の右辺の第２項B(x,y)は、パターンエッジの微小な位置ずれやパターン形状の微小な差異、パターン歪みを許容するための項である。 A(x,y)を求める(数２４)とB(x,y)を求める(数２５)を対比させれば明らかなように、B(x,y)はa a、 b b による、差画像の階調値の変化予測の絶対値である。 A(x,y)によって既知の位置ずれ（とみなす）がキャンセルされるとするならば、 A(x,y)にB(x,y)を加算するのは、位置の合った状態からさらにｘ方向にa a、ｙ方向にb bだけ位置をずらす（とみなす）ことを意味している。つまり、ｘ方向にa a、ｙ方向にb bの位置ずれを許容するのがB(x,y)である。
【０１５４】
また、 A(x,y)からB(x,y)を減ずるのは、位置の合った状態からさらにｘ方向に- a a、ｙ方向に-b bだけ位置をずらすことを意味している。 ｘ方向に- a a、ｙ方向に-b bの位置ずれを許容するのが -B(x,y)である。しきい値に上限、下限を設けることによって、± a a、± b bの位置ずれを許容することになるわけである。 そして、パラメータa a 、 b bの値を適切な値に設定することによって、許容する位置ずれ量を自由にコントロールすることが可能である。
【０１５５】
しきい値を算出する(数２２)及び(数２３)の第３項C(x,y)は、階調値の微小な差異を許容するための項である。 C(x,y)の加算は、g1の階調値がf1の階調値よりもC(x,y)だけ大きいのを許容することを意味し、 C(x,y)の減算は、g1の階調値がf1の階調値よりもC(x,y)だけ小さいのを許容することを意味する。ここでは、 C(x,y)を、局所領域での階調値の代表値（ここではmax値）に比例定数γを掛けた値と、定数εとの和としたが、この関数にこだわる必要はなく、階調値の変動の仕方が既知であれば、それに合った関数するのがよい。例えば、階調値の平方根に変動幅が比例すると分かっていれば、(数２６)の代わりに、 C(x,y)=(max1+max2)1/2*γ＋εとすべきである。 B(x,y)と同様、パラメータγ、εによって、許容する階調値の差異を自由にコントロールすることが可能である。
【０１５６】
欠陥判定回路５０では、差分抽出回路４９の出力sub(x,y)および、しきい値演算回路４８の出力thL(x,y)、thH(x,y)を用いて、
【０１５７】
【数２８】

【０１５８】
を満たせば、位置(x,y)の画素は非欠陥候補、満たさなければ位置(x,y)の画素は欠陥候補と判定する。欠陥判定回路５０は、非欠陥候補画素は0、欠陥候補画素は1以上の値を持つdef(x,y)を出力する。
【０１５９】
特徴抽出回路５０ａでは、ノイズ除去処理（例えば、def(x,y)に対して縮小・膨張の処理を行う）によってノイズ的な出力を削除したあと、近隣の欠陥候補部を一つにまとめる欠陥候補部のマージ処理を行う。その後、一まとまりごとに、重心座標、ＸＹ投影長、面積などの特徴量を算出する。
【０１６０】
全体制御部１０４では、欠陥部の座標を試料上の座標系に変換し、疑似欠陥の削除を行い、最終的に、試料上での位置と特徴量とからなる欠陥データをまとめる。
【０１６１】
この欠陥データは、前記下実施の形態１と同様に、デイスプレイ上に表示することもできるし、出力手段により出力することもできる。
【０１６２】
また、このデイスプレイ上に、階調変換する前の画像または画像のデータと階調変換後の画像または画像のデータとをそれぞれ表示することにより、または、階調変換後の画像または画像のデータを表示することにより、階調変換が正しくなされて検査が行われていることを、作業者が目視で確認することができる。
【０１６３】
本実施の形態によれば、小領域全体としての位置ずれや、個々のパターンエッジの微小な位置ずれや、階調値の微小な差異が許容されるため、正常部を欠陥と誤認識することがなくなる。また、パラメータa a 、 b b、γ、εによって、位置ずれ、階調値の変動の許容量のコントロールを容易に行うことが可能となる。
【０１６４】
〔実施の形態 ３〕
本発明に係るパターン検査方法および装置の第３の実施の形態を図２３に示す。図２３において、図１及び図２と同様に、１はイメージセンサであり、被検査パターンである半導体ウエハ４からの反射光の明るさ、すなわち濃淡に応じた濃淡画像信号を出力するものであり、２はイメージセンサ１から得られる濃淡画像信号をディジタル画像信号９に変換するＡ／Ｄ変換器、３は濃淡画像信号を遅延させる遅延メモリ、４は被検査パターンのある半導体ウエハ、５は被検査パターンの半導体ウエハ４を載置するＸ方向とＹ方向とＺ方向とθ方向（回転）の移動するステージ、６は半導体ウエハ４に対する対物レンズ、７は被検査パターンの半導体ウエハ４を照明する照明光源、８は照明光を反射して対物レンズ６を通して半導体ウエハ４に照射すると共に、半導体ウエハ４からの反射光を透過するハーフミラー、９は濃淡画像信号がＡ／Ｄ変換器で変換されたディジタル画像信号である。このようにして、照明光源７からの照明光を反射させて対物レンズ６を通して半導体ウエハ４に対して、例えば明視野照明を施すように構成している。
【０１６５】
また、３は、画像信号９を繰り返される１セルまたは複数セルピッチ分を記憶して遅延させる遅延メモリであってもよいし、 画像信号９を繰り返される１チップまたは複数チップピッチ分記憶して遅延させる遅延メモリでもよい。
【０１６６】
１１はディジタル画像信号９及び遅延されたディジタル画像信号１０を位置合わせするものであり、ここでは、画素単位で濃淡差が最小となる位置ずれ量を正規化相関により検出し、この位置ずれ量に基づき一方の画像をシフトして、２枚の画像を位置合わせするものである。なお、正規化したのは、位置合せすべき画像間の明るさの違いの影響を軽減するためである。
【０１６７】
即ち、検出画像f(x,y)に対して記憶画像g(x,y)を移動し、相関値が最大となる位置を下式により求める。
【０１６８】
【数２９】

【０１６９】
【数３０】

【０１７０】
【数３１】

【０１７１】
【数３２】

【０１７２】
【数３３】

【０１７３】
ここで、画像はイメージセンサにより連続的に検出されるが、画像を後述するライン単位ごとに分割し、この単位で位置合わせを行う。上式では、検出画像はＸ×Ｙ画素の寸法である。
なお、図示していないが、画像の位置ずれを求める上記した正規化相関は、すべての画像を相手にして行う必要はなく、例えば画像をイメージセンサの長手方向にK分割し、分割した各小画像（X／K×Y画素の大きさ）のうち、情報がある小画像について、行ってもよい。情報があるかどうかの判断は、例えば各小画像を微分し、エッジの有無を検出し、エッジが多い小画像を選ぶ。たとえば、イメージセンサがマルチタップ構成の並列出力可能なリニアイメージセンサの場合、各タップ出力画像が、小画像に相当する。この考え方は、並列出力される画像は、位置ずれが等しいということに基づいている。
【０１７４】
１３は、明るさの異なる画像信号を、明るさを一致させるべく、双方の画像信号の階調を変換する階調変換部である。ここでは、個々の画素毎にゲインとオフセットにより線形変換を実施して、明るさを一致させている。
【０１７５】
【数３４】

【０１７６】
【数３５】

【０１７７】
【数３６】

【０１７８】
【数３７】

【０１７９】
１２は、明るさの異なる画像信号を、明るさを一致させるべく、双方の画像信号を変換する明るさ変換部である。ここでは、画像全体に一括したフィルタ操作を実施して、明るさを一致させている。
【０１８０】
そして、得られた画像信号を比較部１４において比較し、不一致を欠陥として検出するものである。
【０１８１】
２３は、比較する２枚の画像を入力する画像入力部であり、この画像より、散布図作成部２４において散布図を求める。散布図は、縦軸と横軸が、２枚の画像の明るさを示している。２５は、得られた散布図を表示する表示部である。１５は、入力手段であるが、例えば、差画像の絶対値を２値化するしきい値を入力し、入力したしきい値の線分を散布図上にプロットする。この散布図を見れば、入力したしきい値の妥当性が判断しやすい。また、表示された散布図を参照して、画像に適したしきい値を決めることもできる。散布図の１例を図３４に示す。
【０１８２】
Ｗ（ｘ，ｙ，ｄｘ，ｄｙ）＝１のときは以下のように示される。
【０１８３】
【数３８】

【０１８４】
【数３９】

【０１８５】
また、散布図において、プロットしたデータ群に線分を最小２乗法等であてはめ、この線分からのずれである誤差を求めることができる。
【０１８６】
直線を、Ｙ＝ｍ・ｆ（ｘ，ｙ）＋ｎとすると、最小２乗法（ｍ，ｎ）は以下の式にて算出)により直線近似を行う。
【０１８７】
【数４０】

【０１８８】
【数４１】

【０１８９】
直線からの誤差は、例えば下式で求められる。
【０１９０】
【数４２】

【０１９１】
【数４３】

【０１９２】
この誤差に基づき、しきい値を求め、求めたしきい値を散布図上にプロットできる。たとえば、しきい値は、このVeの平方根に比例した値とする。図２４に、これを実現する構成例を示す。
【０１９３】
２６は統計量算出部を示し、上記した線分の当てはめと、線分からの誤差を算出するものである。２７は、この得られた統計量よりしきい値を算出するしきい値算出部である。もちろん、しきい値をユーザ入力可能な構成にしても問題ない。
【０１９４】
散布図に使用する画像は、比較する２枚の画像であり、例えば画素単位の位置合せ後の画像である。画像処理の各段階で、２枚の画像を画像入力部２３に入力可能である。
【０１９５】
図２６及び図２７は、図２３に示した方式に基づき、２枚の画像を処理した例を示したものである。対象は、ライン＆スペースのパターンが、画像の右下部に検出されたものである。左上は、パターンがない領域である。各処理途中での画像のヒストグラム、差の画像の統計量も併せて示している。ヒストグラムからわかるように、最初の段階では、２枚の画像の明るさは一致していない。まず、これを画像を正規化相関により相関値を求め、この相関値が高い位置を求めることにより、画素の単位で位置合せする。次に、位置合せされた２枚の画像について局所階調変換である局所的明るさ補正を実施する。そして、最後に２枚の画像の明るさを一致させるフィルタリングを施し、画像の明るさの一致度をさらに向上させている。
【０１９６】
図２８〜３０は、各処理段階での画像の散布図を示している。画素の単位で位置合せされた段階では、２枚の画像の明るさが一致していないため、散布図において斜め４５度の直線にのらず、直線からのばらつきがみられる。しかし、本発明による局所階調変換である局所的明るさ補正とフィルタリングの処理の後では、散布図が直線に近いところに分布しており、２枚の画像の明るさをそろえる意味で効果があることがわかる。なお、傾きと切片とあるのは、散布図データにフィッティングした線分の傾きと切片である。2枚の画像の一致度の尺度である傾きは、最初0.705であったものが、局所階調変換である局所的明るさ補正後に、0.986となり、さらにフィルタリングの処理の後で0.991となり、明るさの一致度が向上していることがわかる。
【０１９７】
さらに、２枚の画像の一致度を表わす、前述のVeの値も、最初は40.02あったものが、局所階調変換である局所的明るさ補正後に、8.598となり、さらにフィルタリングの処理の後で7.477となり、明るさの一致度が向上していることがわかる。また、Veの値の画像は、Veの値を画像全体ではなく、例えば各画素の周りを含む７×７画素の各領域毎に、直線近似して誤差Veを求めたものを画像として表示したものを、図２８〜図３０は示している。この画像を見れば、どの部分で明るさ合せ誤差が大きいかがわかる。
【０１９８】
また、図３１〜図３３は、画像の局所コントラストに関する散布図を示したものである。ここで、コントラストは、各画素の周囲、例えば３×３画素について最大値と最小値を検出し、最大値−最小値を示す。局所コントラストも、本発明による局所階調変換とフィルタリングの処理の後では、散布図が直線に近いところに分布していることから、これらをそろえる意味もあると考えられる。傾きと切片等は、前述の図と同様の意味をもつ。なお、Veの値の画像は、局所コントラスト画像において、各画素の周りを含む７×７画素の領域に関して直線近似してVeを求めたものである。
【０１９９】
図３４〜図３７は、散布図としきい値の例を示したものである。図３４は、2枚の画像が異なるため、これを誤検出しないように、しきい値が大きく設定されている。図３５は、本発明による局所階調変換である明るさ補正後の散布図であり、2枚の画像の一致度が高いため、設定されたしきい値が小さい。図３６は、明るさ合わせ後の散布図であり、さらにしきい値が小さくなっている。図３７は、片方の画像を、各画素ではなく画像の単位で、線形の階調変換を施したもので、散布図上ではしきい値がオフセットをもっている。
【０２００】
図３８は、画像単位の階調変換を区分的な線形変換とした例を示したものである。この例では、２つの区分に分けている。
【０２０１】
このように、散布図としきい値は、欠陥検出感度等を決める基準或いは、設定したしきい値の妥当性確認など、はばひろく活用できるものである。
【０２０２】
これらの散布図作成、表示、或いは散布図のデータを用いたしきい値算出等は、検査開始前に、画像を検出し、この画像を用いて実施することも可能である。さらに、画像検出に同期して、散布図作成、しきい値設定等を画像毎に行えば、さらに高感度な検査が実現できることは、明白であろう。画像の検出は、各処理が終わった段階で行えばよい。なお、上記したように、画像処理はパイプライン型の処理で実現しているが、そうでない構成のものでも適用できるものである。
【０２０３】
図４１に、別の処理結果を示す。図４０では、まず画像のマッチング(正規化相関)による位置合わせを行う。次に、明るさ補正を行う。明るさ補正では、前述の明るさフィルタリングと局所的明るさ補正を同時に行う。これによれば、より構成が簡単になる。
【０２０４】
なお、上記した例では、リアルタイムで（画像検出と同期して）、係数ｍ、ｎ等を求めて、画像の明るさを補正した。その簡略版として、各画像の明るさについて、係数ｍ、ｎを求めておき、これを参照する形で、明るさ補正を行う手段も考えられる。各画像とは、比較する２枚の画像でもよい。即ち、比較する２枚の画像の明るさを入力し、これに対する係数ｍ、ｎを求めておく。そして、作成された参照表をもとに明るさ補正を行うのである。
【０２０５】
次に、画像の大きさに関して説明する。画像の大きさ、即ち画像の位置合せ（マッチング）の単位は、次の方法で決定できる。まず、比較する２枚の画像の位置ずれ量を細かく分割した単位で求める。これを図３９に示す。Ｘ方向とＹ方向に分離して検出している。この位置ずれデータをスペクトル分析すると、図４０のような波形が得られる。スペクトル分析図では、縦軸はスペクトル密度、横軸は周波数を示す。この図において、最も高い周波数であり、かつ密度の高い周波数に着目する。この図の場合、０．０１１となる。この周波数は、例えばステージの走行特性等の装置特性、振動特性により決まるものである。
【０２０６】
スペクトル分析結果は、２枚の画像の位置ずれがこの周波数で繰り返していることを表わしており、この逆数である８８ラインを画像の単位、即ちマッチングの単位とすると、画像内に位置ずれのピークｔｏピークが現れ、位置ずれ量が大きい場合精度の高い位置合せが困難となる。しかし、画像の単位をこの周波数の逆数の１／４にすると、ピークの位置ずれの１／２以下に位置ずれ量を低減することができる。さらに、周波数の逆数の１／８にすると、ピークの位置ずれの１／４以下に位置ずれ量を低減することができる。
【０２０７】
このように、画像の単位を細かくすればするほど、画像の位置合せの精度を向上できるはずであるが、画像内に含まれるべきパターンの情報が少なくなるので、結果としては画像の位置合せ精度が上がらない。従って、スペクトル分析結果からは、必要とする位置合せ精度をもとに、画像の大きさに制限がなされ、パターン情報の確保の観点からは、比較するパターンに応じてであるがパターンのスペース情報（パターンが形成されていない領域に関する情報）をもとに画像の大きさの下限が決定できる。図２５に、この構成を示す。なお、上記した説明では、最も高い周波数に着目したが、位置ずれの量に着目し、その大きなものに相当する周波数に着目しても効果的である。
【０２０８】
なお、上記は、ｘ、ｙ成分に分離して上記を行うこともできるし、蓄積型のリニアイメージセンサの場合のように、ステージの進行方向のみに着目して実施してもよい。
【０２０９】
上記発明の実施の形態１、２、３によれば、場所によるパターンの明るさの違いに影響されることなく、欠陥を高感度に検出することができる。また、メモリマット部２１など暗い領域において、明るさが大きくばらつくパターンにおいても高感度に検査できる。また、装置が有する振動特性等に影響を受けることなく、高精度な画像のマッチングが実現できる。従って、従来に比べ、信頼性の高い検査を実現することができる。
【０２１０】
以上、本発明に係る実施の形態について、主に光学顕微鏡や走査型電子顕微鏡を用いた比較検査方法について述べたが、他の赤外線やＸ線により得られた画像検出に用いた場合にも、同様に有効であることはいうまでもない。
【０２１１】
【発明の効果】
本発明の構成によれば、多層のパターンの膜厚の違いによって生ずる明るさの違いによって全体の検査感度が律束されることなく、欠陥を高感度に検出することができる効果を奏する。
【０２１２】
また、高精度の画像マッチングの実現に効果を奏する。さらに、比較すべき２つの検出画像信号の明るさや微分等の処理を施した明るさ情報、或いは指定した画像処理途中の明るさ情報を、縦軸と横軸とした散布図を併せて使うことにより、しきい値の決定やその妥当性確認が有効にできる。
【図面の簡単な説明】
【図１】本発明の一実施形態に係る被検査パターンの欠陥検査装置の構成図である。
【図２】本発明の一実施形態に係る被検査パターンの欠陥検査装置の構成図である。
【図３】画像明るさ一致フィルタ操作部の説明図である。
【図４】階調変換部の説明図である。
【図５】双対フィルタの例を示す図である。
【図６】本発明にかかる検出画像と差画像の例を示す図である。
【図７】本発明の階調変換を実施した例を示す図である。
【図８】本発明の階調変換を実施した例を示す図である。
【図９】本発明の階調変換を実施した例を示す図である。
【図１０】本発明の階調変換を実施した例を示す図である。
【図１１】本発明の階調変換を実施した例を示す図である。
【図１２】本発明のしきい値設定方式の説明図である。
【図１３】本発明のしきい値設定方式の説明図である。
【図１４】本発明の画像明るさ一致フィルタ操作部の動作説明図である。
【図１５】被検査パターンのメモリチップにおけるメモリマット部と周辺回路部の略示説明図である。
【図１６】図１０のメモリチップにおけるメモリマット部と周辺回路部における明るさのヒストグラムである。
【図１７】ＣＭＰプロセス処理のフローを説明する図である。
【図１８】ＣＭＰ処理された、異なるメモリチップにおけるメモリマット部と周辺回路部における明るさのヒストグラムである。
【図１９】本発明の一実施形態に係る被検査パターンの欠陥検査装置の構成図である。
【図２０】しきい値演算回路の構成を説明する図である。
【図２１】２枚の画像の差の２乗和を表す図である。
【図２２】２枚の画像の差の絶対値の和を表す図である。
【図２３】本発明の一実施形態に係る被検査パターンの欠陥検査装置の構成図である。
【図２４】本発明の一実施形態に係る散布図作成、表示を説明する図である。
【図２５】本発明の一実施形態に係る被検査パターンの欠陥検査装置の構成図である。
【図２６】比較する２枚の画像に関して、各画像処理の段階での結果を表わす図である。
【図２７】比較する２枚の画像に関して、各画像処理の段階での結果を表わす図である。
【図２８】比較する２枚の画像に関して、各画像処理の段階での散布図を表わす図である。
【図２９】比較する２枚の画像に関して、各画像処理の段階での散布図を表わす図である。
【図３０】比較する２枚の画像に関して、各画像処理の段階での散布図を表わす図である。
【図３１】比較する２枚の画像に関して、各画像処理の段階での局所コントラストの散布図を表わす図である。
【図３２】比較する２枚の画像に関して、各画像処理の段階での局所コントラストの散布図を表わす図である。
【図３３】比較する２枚の画像に関して、各画像処理の段階での局所コントラストの散布図を表わす図である。
【図３４】散布図の１例を表わす図である。
【図３５】散布図の１例を表わす図である。
【図３６】散布図の１例を表わす図である。
【図３７】散布図の１例を表わす図である。
【図３８】散布図の１例を表わす図である。
【図３９】画像の位置ずれ量を説明する図である。
【図４０】スペクトル分析を説明する図である。
【図４１】比較する２枚の画像に関して、各画像処理の段階での結果を表わす図である。
【符号の説明】
１…イメージセンサ、２…Ａ／Ｄ変換器、３…遅延メモリ、４…半導体ウエハ、５…Ｘ、Ｙ、Ｚ、θステージ、６…対物レンズ、７…照明光源、８…ハーフミラ、９…画像信号、１０…画像信号、１１…画素単位位置合わせ部、１２…画像明るさ一致フィルタ操作部、１３…階調変換部、１４…比較部、１５…入力手段、１６…ＣＰＵ、１７…記憶装置、２０…チップメモリ、２１…メモリマット部、２２…周辺回路部（非繰り返しパターンであるが、複数の繰り返しピッチをもつ繰り返しパターンを含む）、２３…画像入力部、２４…散布図作成部、２５…表示部、２６…統計量算出部、２７…しきい値算出部（或いは入力部）、３１…電子銃、３２…磁界レンズ、３３…対物レンズ、３４…偏向器、３５…検出器、３６…アンプ、３７…グランド電極、３８…リターデイング電極、３９…Ａ／Ｄ変換器、４０…前処理回路、４１…遅延回路、４２…画素単位の位置合わせ部、４６…局所階調変換回路、４８…しきい値演算回路、４９…差分抽出回路、５０…欠陥判定回路、１０１…検出部、１０２…画像取り出し部、１０３…画像処理部、１０４…全体制御部。[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an appearance inspection for detecting a defect in a pattern to be inspected, and more particularly to a method and apparatus for inspecting a defect in a pattern to be inspected in a semiconductor wafer or a liquid crystal display.
[0002]
[Prior art]
Conventionally, this type of inspection apparatus detects an image of a pattern to be inspected by an image sensor such as a line sensor while moving the pattern to be inspected as in the technique described in Japanese Patent Application Laid-Open No. 55-74409. The inconsistency is recognized as a defect by comparing the density of the image signal and the image signal delayed by a predetermined time.
[0003]
Further, as described in Japanese Patent Publication No. 8-10463, the comparison was performed by aligning two images.
[0004]
The conventional defect recognition method will be described in detail with reference to FIGS. FIG. 15 is a schematic explanatory diagram of a memory mat portion and a peripheral circuit portion in a memory chip having a conventional pattern to be inspected. FIG. 16 is a brightness histogram of the memory mat portion and the peripheral circuit portion in the memory chip of FIG. FIG. 17 is a schematic view of a pattern to be inspected that has been flattened by CMP (chemical mechanical) or the like.
[0005]
As shown in FIG. 15, a large number of memory chips 20 are arranged on a semiconductor wafer. The memory chip 20 can be roughly divided into a memory mat portion 21 and a peripheral circuit portion 22. The memory mat portion 21 is a set of small repetitive patterns (cells), and the peripheral circuit portion 22 is basically a set of random patterns. However, in many cases, it can be regarded as an aggregate of repetitive patterns having a plurality of different cell pitches when viewed in detail.
[0006]
FIG. 16 shows the brightness distribution in the memory mat portion 21 and the peripheral circuit portion 22 of FIG. 15, that is, the frequency (histogram) with respect to the brightness in the memory chip as a maximum of 1024 gradations in the 10-bit configuration. However, the memory mat portion 21 has a high pattern density and is generally dark. On the other hand, the peripheral circuit portion 22 has a low pattern density and is generally bright.
[0007]
[Problems to be solved by the invention]
In the planarization process such as CMP shown in FIG. 17, as can be seen from the histogram shown in FIG. 18, the circuit pattern in the memory mat section 21 causes a difference in brightness due to a difference in pattern film thickness. . In this figure, the wiring layer is deposited and then flattened by CMP processing. In such a pattern, the film thickness locally fluctuates and uneven brightness tends to occur. In the case of such a pattern, the brightness of the patterns shown in FIG. 16 and FIG. 17 is compared, and if a threshold value is set so as not to erroneously detect a difference in brightness, the defect detection sensitivity is extremely lowered. . Such a difference in brightness can be offset to some extent by using illumination light having a wide wavelength band, but there is a limit to the pattern subjected to the CMP process because the variation in brightness may be large. For this reason, it has been desired to detect minute defects from patterns with different brightness.
[0008]
Further, conventionally, the sum of squares of the difference between two images is calculated, and a paraboloid is applied to the difference to detect the image misalignment. However, in this case, there is no guarantee that the two images to be compared match, and there has been a demand for optimal matching for comparison. In FIG. 21, the sum of squares of differences for each pixel of two images (f (x, y) in FIG. 6 described later) is shifted in the range of ± 1 pixel in the x and y directions. This is the result of the experiment. The horizontal axis is the x direction, and the vertical axis is the y direction. The value in the figure is the sum of squared differences. Here, the same image (f (x, y) in FIG. 6) is used. That is, Σ (f (x, y) −f (x ± 1, y ± 1)) 2 is calculated as the sum of squares of the differences. As can be seen from FIG. 21, even in the same image, the sum of squares of the difference is not symmetric about (0, 0) but has an asymmetry of about 0.6%. Since they are the same image, the square sum of the difference is 0 at (0, 0). Accordingly, even if a parabola is applied to this data and a position where the sum of squares of the difference is minimized with a resolution equal to or smaller than the pixel size is obtained, an accurate positional shift (here, (0, 0)) cannot be detected.
[0009]
Furthermore, the brightness is different in a wafer of a planarization process such as CMP. The influence of this brightness difference is shown. Here, a single image and its brightness multiplied by 1.1 were used. The 1.1 times is normal or small as the CMP wafer brightness fluctuation. The experimental results are shown in FIG. The sum of absolute values of differences is displayed. The minimum position is (0, 1). There is a large error at the pixel level rather than the resolution below the pixel. The sum of squares of the differences has the same tendency, but it can be seen from these data that the image misalignment cannot be obtained accurately. Of course, the same tendency is observed even at 1.05 times. Thus, it is necessary to say that a method having an extremely large error is to apply a paraboloid to the sum of squares of the difference and obtain a minimum position from this.
[0010]
An object of the present invention is to solve the above-described conventional technical problem, and allows a pattern having different brightness to be comparatively inspected, and is a pattern to be inspected that can always inspect defects with high sensitivity and high reliability. It is to provide a defect inspection method.
[0011]
Another object of the present invention is to provide a defect inspection method for a pattern to be inspected using a highly accurate image matching method.
[0012]
It is another object of the present invention to provide a highly sensitive defect detection method even when a planarized wafer pattern such as CMP is targeted.
[0013]
[Means for Solving the Problems]
  In order to achieve the above object, according to the present invention, in a method for inspecting a defect of a plurality of sets of patterns originally formed on a substrate so as to be the same, a first pattern to be inspected is detected and this first pattern is detected. The first image of the pattern to be inspected is obtained, the first image is stored, the second pattern to be inspected is detected to obtain the second image of the second pattern to be inspected, and the stored first And the second imageThe first image and the second image obtained by performing the adjustment of the brightness of the entire image and the adjustment of the local brightness, and performing the adjustment of the brightness of the entire image and the adjustment of the local brightness. The pattern is inspected by comparing the first image with the second image, and adjusting the brightness of the entire image of the first image and the second image stored together with each other. By performing a batch filter operation on the entire image withI did it.
[0014]
  Further, according to the present invention, the brightness of the whole of the first image and the second image stored is adjusted so that the square error of the density between the stored first image and the second image is Operation is performed by using a filter that minimizes the brightness, and the brightness of the entire image is adjusted so that the square error in brightness between the first image and the second image is minimized. .
[0015]
  Further, according to the present invention, the first image and the second image can be obtained by linearly adjusting the brightness of the stored first image and the second image by linear transformation having a gain and an offset. Is converted so that the brightness of the pixels is almost the same, and weighting is performed according to the difference in brightness between the pixel of interest and the surrounding pixels in the region to be locally matched to the linear conversion having the gain and the offset. I added.
[0016]
  Then, the stored brightness of the first image and the second image is locally adjusted before the pixel-by-pixel alignment of the two images.
[0017]
  Alternatively, the brightness of the stored first image and second image is locally adjusted after the pixel unit alignment of the two images.
[0018]
  According to the present invention, in a method for inspecting a defect of a plurality of sets of patterns originally formed to be the same on a substrate, the first inspected pattern is imaged to capture the first inspected pattern. A first image is obtained, the first image is stored, a second pattern to be inspected is imaged to obtain a second image of the second pattern to be inspected, and the stored first image and second The positional deviation between the first image and the second image corrected for the positional deviation is detected, and the local brightness difference between the first image and the second image is corrected, and the local difference is detected according to the detected brightness difference. The local brightness is adjusted by performing appropriate gradation conversion, and the brightness of the entire first image and the entire second image are adjusted to match this local brightness. The pattern is inspected by comparing the first and second images combined And adjusting the brightness of the whole of the first image and the whole of the second image to minimize the square error of the light and shade of the whole of the first image and the whole of the second image. Thus, the filtering is performed on the entire first image and the entire second image.
[0019]
  Further, according to the present invention, the brightness of the first image and the second image is adjusted by linearly transforming the brightness of the first image and the second image locally with a gain and an offset. Was converted to be almost the same.
[0020]
  According to the present invention, in a defect inspection apparatus for inspecting defects of a plurality of sets of patterns originally formed on the substrate so as to be the same, the first inspection pattern is imaged and the first inspection target is imaged. An imaging unit that obtains a first image of a pattern and captures a second pattern to be inspected to obtain a second image of the second pattern to be inspected, and a first image obtained by imaging with the imaging unit Storage means for storing, gradation conversion means for locally gradation-converting at least one of the first image and the second image, and at least one of them locally by the gradation conversion means A brightness correction unit that matches the brightness of the first image and the second image that have undergone gradation conversion, and a comparison between the first image and the second image that match the brightness by the brightness correction unit And a defect detection means for detecting a defect in the pattern. Form was.
[0021]
  The gradation converting means is configured to convert the brightness of the first image and the second image to be compared by linear conversion having a gain and an offset.
[0022]
  In addition, a misalignment correction unit that corrects misalignment between the first image and the second image is provided, and the gradation conversion unit includes the first image and the second image in which the misalignment is corrected by the misalignment correction unit. And at least one of them is locally gradation-converted.
[0033]
DETAILED DESCRIPTION OF THE INVENTION
Each embodiment of the present invention will be described with reference to FIGS. 1 to 14 and FIGS. 23 to 30.
[0034]
[Embodiment 1]
An inspection pattern defect inspection method and apparatus according to the present invention will be described. 1 and 2 are block diagrams of a defect inspection apparatus for a pattern to be inspected according to an embodiment of the present invention.
[0035]
In the present embodiment, a semiconductor wafer inspection pattern will be described as an example.
[0036]
1 and 2, reference numeral 1 denotes an image sensor, which outputs a grayscale image signal corresponding to the brightness of reflected light from the semiconductor wafer 4 that is a pattern to be inspected, that is, grayscale, and 2 is an image sensor 1. An A / D converter for converting the grayscale image signal obtained from the digital image signal 9 into a digital image signal 9, 3 a delay memory for delaying the grayscale image signal, 4 a semiconductor wafer having a pattern to be inspected, and 5 a semiconductor wafer 4 having a pattern to be inspected A stage that moves in the X, Y, Z, and θ directions (rotation), 6 is an objective lens for the semiconductor wafer 4, 7 is an illumination light source that illuminates the semiconductor wafer 4 of the pattern to be inspected, and 8 is illumination. A half mirror that reflects light and irradiates the semiconductor wafer 4 through the objective lens 6 and transmits the reflected light from the semiconductor wafer 4, and 9 indicates that the grayscale image signal is A / D It is converted digital image signal in exchanger. In this way, the illumination light from the illumination light source 7 is reflected, and the semiconductor wafer 4 is configured to perform, for example, bright field illumination through the objective lens 6.
[0037]
Further, 3 may be a delay memory that stores and delays the image signal 9 for one or a plurality of repeated cell pitches, or stores and delays the image signal 9 for one or more repeated chip pitches. A delay memory may be used.
[0038]
11 is used to align the digital image signal 9 and the delayed digital image signal 10, and here, a misregistration amount that minimizes the shading difference is detected in units of pixels, and one image is based on the misregistration amount. To align the two images. Here, the image is continuously detected by the image sensor, but the image is divided into, for example, every 256 lines (the number of lines is determined by a method described later), and alignment is performed in this unit. Reference numeral 12 denotes a brightness conversion unit that converts image signals having different brightnesses so that the brightness of the image signals is matched. Here, filter operations are performed collectively on the entire image to match the brightness.
[0039]
A gradation conversion unit 13 converts the gradations of the image signals having different brightnesses so that the brightnesses of the image signals match. Here, linear transformation is performed for each individual pixel using a gain and an offset, and the brightness is matched. Then, the obtained image signals are compared in the comparison unit 14 to detect a mismatch as a defect.
[0040]
The detected image signal is sequentially subjected to a certain process by pipeline type image processing, and finally a defect and its feature are output.
[0041]
In the above example, bright field illumination is used as illumination, but the illumination is not limited to this, and any illumination that can be used as microscope illumination such as dark field illumination or annular illumination may be used. Needless to say, it can also be applied to illumination with an electron beam.
[0042]
The lighting conditions may be variously changed, the inspection is performed a plurality of times, and the logical sum of the inspection results of the plurality of times may be taken as a final result. Alternatively, a logical product may be taken and identified as a defect, and the process diagnosis may be made based on the defect distribution and the number of defects, for example. In this case, the review for visually confirming the inconsistent portion is unnecessary, and the work can be simplified and simplified.
[0043]
Next, the operation of the inspection apparatus having the above configuration will be described with reference to FIGS. In FIG. 1 and FIG. 2, the processing order is changed.
[0044]
In FIG. 1 and FIG. 2, the stage 5 is scanned in the X direction with illumination light converged by the objective lens 6 and moved at a constant speed with respect to the target region of the semiconductor wafer 4 of the pattern to be inspected, while the image sensor 1 A pattern to be inspected formed on the semiconductor wafer 4, that is, brightness information (grayscale image signal) of the memory mat portion 21 and the peripheral circuit portion 22 in the chip 20 is detected.
[0045]
When the movement for one row is completed, it moves to the next row at a high speed in the Y direction and is positioned. That is, the inspection is performed by repeating the constant speed movement and the high speed movement. Of course, step-and-repeat inspection is also acceptable. The A / D converter 2 converts the output (grayscale image signal) of the image sensor 1 into a digital image signal 9. The digital image signal 9 has a 10-bit configuration. Of course, if there are about 6 bits, there is no particular problem in image processing, but a certain number of bits are required to detect minute defects.
[0046]
First, image alignment in pixel units will be described. Image alignment in pixel units is performed by calculating the density difference (difference between the value of each pixel of the image and the value of the corresponding pixel) while shifting one of the two images to be compared by the unit of pixel, thereby minimizing the density difference. The amount of displacement is obtained. The range of image misregistration detection is, for example, a maximum of ± 3 pixels, and is variable according to the pattern design rules. The two images are aligned by shifting the position of one of the images by the amount of displacement obtained.
[0047]
This will be described using the following.
[0048]
[Expression 1]

[0049]
The position shift detection is to detect Δx and Δy where S (Δx, Δy) is min.
[0050]
However, since the minimum position can be obtained only in units of pixels, it is added as an offset depending on whether the true position is closer to the obtained Δx or Δy.
[0051]
Based on the following formula, 1 is added to Δx, Δy or left as it is.
[0052]
That is,
[0053]
[Expression 2]

[0054]
[Equation 3]

[0055]
[Expression 4]

[0056]
[Equation 5]

[0057]
Δx ++ means Δx = Δx + 1.
[0058]
In this way, the two images are aligned by shifting the position of one image by the amount of positional deviation always obtained by the alignment. That is, the image f is always moved to the upper right to obtain a new image f ′, and the movement direction can be specified from one of four types (lower right movement, upper left movement, lower left movement, upper right movement). . This leads to hardware simplification.
[0059]
FIG. 3 shows the brightness matching filter operation unit 12 in detail. For the two images f (x, y) and g (x, y) aligned in pixel units, first, filters F and F ′ that minimize the following values in the image are obtained.
[0060]
[Formula 6]

[0061]
The filters F and F ′ have a size of 2 × 2 pixels, for example.
[0062]
FIG. 5 shows an example of the filter. The filters F and F ′ are symmetric and dual as shown in FIG. With such a dual form, the coefficient of the parameter of the filter can be solved using the least square method.
[0063]
[Expression 7]

[0064]
[Equation 8]

[0065]
However,
[0066]
[Equation 9]

[0067]
[Expression 10]

[0068]
## EQU11 ##

[0069]
In this method, filtering is performed on two target images, and matching is performed by minimizing the square error of the density. It does not require repetitive calculations and can be realized by a single calculation.
[0070]
A feature of the above method is that the filter coefficients α and β are obtained so that the shades of two images are well matched with each other in the sense that the square error is minimized. In particular, these parameters do not necessarily represent the true value of the image displacement amount. For example, as described in the prior art, a method is also conceivable in which a paraboloid is applied to S (Δx, Δy), its minimum position is calculated, and then interpolation pixels are obtained by interpolation based on this calculated position. However, in this case, there are no norms or conditions to be satisfied with respect to brightness, and in that sense, it is not guaranteed that the obtained image is used for comparative inspection. Further, when the brightness is different, it is unclear what the calculated displacement represents. Moreover, even if the minimum position calculated by approximating the paraboloid matches the position described in the present embodiment, the images to be compared do not match.
[0071]
The proposed matching method guarantees that the square error of the brightness of two images is minimized, and in this respect, it is different from other methods. As shown in FIG. 14, since it is a linear approximation, the coefficient α of the filter has an error when it is regarded as a positional deviation. However, the brightness obtained is consistent. This method can essentially reduce the difference in shading between images, and is considered to be a very convenient method for comparative inspection.
[0072]
Further, the calculation of the filter coefficients α and β can be performed analytically without repeated calculation, and is suitable for hardware implementation.
FIG. 4 shows the local gradation converting unit 13 in detail. Parameters for minimizing the following values in a predetermined area in an image with respect to two images f (x, y) and g (x, y) that are aligned in pixel units and output as brightness matching filter operations a and b (a: gain, b: equivalent to offset) are obtained.
[0073]
[Expression 12]

[0074]
The parameters a and b are obtained by partially differentiating the above values with a and b and setting them to zero. For example, the defined area is a range of 7 around each point.
[0075]
One g (x, y) of the image signal is calculated using the obtained parameters.
[0076]
[Formula 13]

[0077]
And convert. As a result, images with consistent brightness can be obtained. The parameters a and b can take different values for each position (x, y).
[0078]
[Expression 14]

[0079]
[Expression 15]

[0080]
Here, the number of pixels in the range of Σ is MN.
[0081]
Further, in the Σ range, the brightness of the focused center pixel and the surrounding pixels are compared, and when the brightness is significantly different, it is not necessary to add this.
[0082]
Alternatively, although the addition itself is performed, it is effective to add a weight to lower the contribution rate when adding. For example, when the brightness of the pixel of interest (x, y) is c and the brightness of another pixel in the range of Σ is d, the weight W (x, y) is
[0083]
[Expression 16]

[0084]
And Here, max [] is the maximum value detection, and c and d are both 8-bit gradation brightness. D is a constant. As a result, when the brightness of the focused center pixel and surrounding pixels are similar, the weight is set to a value close to 1, and when the brightness is not similar, the weight is set to a small value. Although D is a constant, it may be varied according to the brightness, and D = func (c). Further, it may be determined by whether or not they belong to the same pattern, and if the average brightness of different patterns is μ, D = | c−μ | may be given. If there are three or more types of patterns, the difference in brightness between similar patterns may be D. Of course, it is not always necessary to stick to such a form, as long as the weight can be appropriately given.
[0085]
FIG. 6 shows an example of two detected images. The detected two images have a great difference in brightness as shown in the figure. The two images were aligned with pixel accuracy and a brightness matching filter operation was performed. However, the difference in brightness is too large in the main image, and as shown in the figure, there is a large mismatch in the difference image. A gradation conversion process was performed on such an image.
[0086]
Examples of processing are shown in FIGS. 7 to 9 show two detected images f (x, y), g (x, y), and a brightness histogram. In addition, an image a * g (x, y) + b after conversion and a brightness histogram are shown. Here, D = 70. As can be seen from the histogram shown in FIG. 7, the value of D is a value corresponding to the difference between the distribution averages of the bimodal histogram distribution. That is, the weight W set with such D is an index as to whether or not they belong to the same distribution. The defined area is a 7 × 7 pixel range around each point. It can be seen from FIGS. 7 to 9 that the brightness histograms become substantially equal by the conversion. Here, in the experimental results for the images of FIGS. 7 to 9, a = 1.41 and b = 0 were obtained at certain points in the image. In this image, it can be seen that the brightness gain is greatly different (41%).
[0087]
From this example, a method in which the offset b is always set to 0 and only the gain is variable can be considered. These may be determined according to the characteristics of the target pattern, the device configuration, and the like.
[0088]
10 and 11 show the difference between images obtained by such conversion. In FIG. 10 and FIG. 11, the first three areas are defined areas of 3 × 3, 5 × 5, and 7 × 7 pixels around each point. Further, the weight at this time is W (x, y) = 1. Further, the last one shows that the weight is in accordance with the above-described W (x, y) for the range of 7 × 7 pixels. From these figures, it can be seen that when the area is small, the brightness is adjusted locally and the discrepancy between the images is reduced. This increases the allowable range of brightness, but misses minute defects. For this reason, it is necessary to expand the area according to the defect to be detected. However, when the weight is fixed to 1, the boundary portion of the pattern area is detected as a mismatch, that is, as a false alarm. In the case of weighting, the influence of the boundary is reduced, the brightness of almost two images is matched, and a minute defect can be detected.
[0089]
The area such as 7 × 7 pixels is not necessarily a square, and may be a rectangle, a polygon, or a circle. Further, instead of such a very small region, it may be a region of several hundreds × several hundred pixels, and the point is only required to be within a range in which variations in brightness can be absorbed.
[0090]
In addition to this, the weighting method can also compare the brightness of the focused center pixel and the surrounding pixels, and set the weight to 0 when the difference in brightness is larger than the threshold value.
[0091]
In addition to the above, the following gradation conversion is also conceivable.
[0092]
[Expression 17]

[0093]
Here, σf and mf are a standard deviation and an average value in an area near the point (x, y) in the image f (x, y), respectively, and σg and mg are respectively the images g (x, y, In y), the standard deviation and the average value in the defined area near the point (x, y).
With the above-described conversion, the brightness of the image g (x, y) can be matched with f (x, y).
[0094]
W (x, y) may be the above-described weight, or may be a correlation coefficient of image data in an area defined by images f (x, y) and g (x, y).
[0095]
This method also has a feature that the histograms of the two images coincide as a result.
[0096]
Both methods are in the form of linear transformation consisting of gain and offset.
[0097]
The above tone conversion is local brightness conversion in the vicinity of the target pixel. Of course, the present invention may be applied to the entire image, that is, the entire 256 lines in accordance with the purpose and image characteristics. Also, which of the two images is to be matched with the brightness of one of the images can be adjusted, for example, by calculating an average brightness value for each image. In addition, it is possible to calculate an average brightness value for each predetermined area or each point, and to adjust to the brighter one.
[0098]
In FIG. 1, tone conversion is performed after the image brightness matching filter operation. However, this order may be changed as shown in FIG.
[0099]
Next, the comparator 14 may be a method developed by the present inventors, such as that shown in the technique described in Japanese Patent Application Laid-Open No. Sho 61-212708, etc., and determines the image difference image detection circuit and the difference image. A mismatch detection circuit that binarizes with a threshold value, and a feature extraction circuit that calculates an area, length (projection length), coordinates, and the like from the binarized output.
[0100]
With reference to FIG. 12 and FIG. 13, the threshold value setting in the binarization according to the present invention will be described.
[0101]
As described above, when the difference image is binarized, false information is likely to occur at the boundary between regions. Therefore, as shown in FIG. 12, the difference between the maximum value and the minimum value in the local region of the image, the average value, and the differential value x or y, whichever is greater (hereinafter referred to as local data) is calculated for each point. . A threshold value is generated by multiplying these values by a separately defined parameter and adding them, so-called product-sum operation. As a result, for example, where the brightness change at the boundary of the region is large, the differential value and the like become large, so that the threshold value becomes large and the occurrence of false information can be prevented in advance. Of course, the larger one of the difference between the maximum value and the minimum value, the average value, and the differential value x or y is not necessarily provided, and any one of them may be provided. For example, when the above tone conversion is performed, there is no need to calculate an average value.
[0102]
If the difference between the images is binarized using such a threshold value, the problem of false alarm can be reduced more effectively. The local data can be determined in size by obtaining a distribution from a scatter diagram described later. 31 to 33 show scatter diagrams regarding the difference between the maximum value and the minimum value in the local region of the image. A line segment is applied to this distribution data, and an error from the line segment is obtained. Such processing is performed on each local data, and the threshold value can be determined by the product sum.
[0103]
For example, as the threshold Th
Th = C3 × local contrast + C2 × brightness average
Think of it. Here, the local contrast image is represented by the maximum value-minimum value of 3 × 3 pixels for each pixel.
The brightness average image is represented by a moving average of 3 × 3 pixels for each pixel.
[0104]
The local contrast images of the two images to be compared are f (x, y) and g (x, y),
[0105]
[Expression 18]

[0106]
Ve obtained from the above equation is σk.
[0107]
Similarly, let f (x, y) and g (x, y) be the average images of brightness, and let Ve be obtained be σa.
[0108]
From these, the following equation (Equation 19) is determined.
[0109]
[Equation 19]

[0110]
The same is done by changing the image, and the coefficients C2 and C3 are obtained.
[0111]
Assuming that Th is selected, in the scatter diagram of the local contrast, the distance from the straight line with the slope 1 (m = 1) and the intercept 0 (n = 0) to each plot data point, and the standard deviation corresponding to the error Find σk. Similarly, in the scatter diagram of the average brightness, the standard deviation σa corresponding to the error, which is the distance from the straight line of slope 1 and intercept 0 to each plot data point, is obtained. Further, in the scatter diagram of the brightness of the original two images, the standard deviation σg corresponding to the error is obtained from the distance from the straight line with the slope 1 and the intercept 0 to each plot data point.
[0112]
By substituting these values into the above equation of Th, equations of C2 and C3 as shown in (Equation 19) are obtained. If this operation is performed for images at different locations, equations for other C2 and C3 with different coefficients are obtained. Therefore, when these are solved as simultaneous equations, the coefficients C2 and C3 are uniquely determined. Thus, the threshold value Th is determined from C2 and C3. Of course, the threshold Th is
Th = C3 × local contrast + C2 × average brightness + offset
It is good.
[0113]
Furthermore, as another setting method, the floating threshold value to be obtained is expressed by (Expression 20) as a linear combination of local contrast of brightness and average value, and using scatter diagram information of two images to be compared. The parameter may be calculated by multiple regression analysis.
[0114]
[Expression 20]

[0115]
The setting procedure at this time is shown below.
[0116]
(1) A plurality of images (a set of two chips) are detected.
[0117]
(2) A brightness scatter diagram is created from the detected image and reference image data. (Use an image that does not include a defect, or use an image that does not include a defect.)
(3) A point that envelops a set of data in the scatter diagram is obtained (a point of frequency 1 is extracted in the evaluation), and local contrast and average value data are extracted from the image pixels corresponding to the point. .
[0118]
(4) Parameters C0 to C4 are adjusted by multiple regression analysis based on the information obtained in (3) above.
[0119]
(5) The data to be used is selected according to the p value (significance level) (a combination in which the p value becomes a sufficiently reliable value (0.05 or less) is obtained).
[0120]
(6) A threshold image is calculated from the obtained parameters C0 to C4 and compared with the difference image.
[0121]
(7) If there is a false report, the parameters C0 to C4 are adjusted in addition to the false report information.
[0122]
(8) Perform a test inspection.
[0123]
(9) When false information is generated, the above (7) and (8) are repeated.
[0124]
Further, as shown in FIG. 13, it is possible to use a lookup table (LUT) instead of a product-sum operation of coefficients and errors as described above. As shown in FIGS. 12 and 13, the local maximum value and the local minimum value of the image are obtained, the contrast is obtained from the difference, and this is input to the LUT. Similarly, the local average value of the image is obtained and input to the LUT. These LUT outputs are input to another LUT, and a threshold value is obtained. In order to make the LUT scale appropriate, the number of bits used in FIGS. 12 and 13 is limited to 8 → 6 bits. The obtained threshold value is input to the comparison circuit (FIGS. 1 and 2) 14 and used as a threshold value for binarization. The data in the LUT is created by using various images and obtaining an error or the like in the same procedure as described above, and interpolating the error or the like.
[0125]
Needless to say, an image that is easily misdetected is suitable for the image to be selected. Conventionally, there is no such determination method, and not only the determination procedure is determined by the present invention, but also a very logical determination. It is a feature.
[0126]
In FIG. 1 and FIG. 2, the CPU 16 inputs the coordinates such as the arrangement data in the chip on the semiconductor wafer 4 obtained based on the design information by the input means 15 constituted by a keyboard, a disk, etc. Then, based on the coordinates such as the array data in the chip on the input semiconductor wafer 4, defect inspection data is created and stored in the storage device 17.
[0127]
The defect inspection data can be displayed on a display means such as a display as needed, or can be output by an output means.
[0128]
Also, on the display, the image or image data before gradation conversion and the image or image data after gradation conversion are displayed respectively, or the image or image data after gradation conversion is displayed. Thus, the operator can visually confirm that the gradation conversion is correctly performed and the inspection is performed.
[0129]
As a result, images can be compared with extremely high accuracy, and the purpose intended by the present invention can be realized with higher sensitivity.
In addition, although the example which used bright field illumination was shown in the present Example, you may make into object the image obtained by dark field illumination. In addition, the types of defects include shape defects such as short and open, and other foreign matters.
[0130]
[Embodiment 2]
A second embodiment of the pattern inspection method and apparatus according to the present invention is shown in FIG. Here, a sample is scanned with an electron beam, electrons generated from the wafer are detected by irradiation of the electron beam, an electron beam image of the scanned part is obtained based on the intensity change, and pattern inspection is performed using the electron beam image. Do. In contrast to the problem to be solved by the above-described invention, in the second embodiment, the threshold value for defect determination is set for each pixel in consideration of the positional deviation of the pattern, the difference in gradation value, and the like. To deal with it.
[0131]
This system includes a detection unit 101, an image extraction unit 102, an image processing unit 103, and an overall control unit 104 that controls the entire system.
[0132]
First, the detection unit 101 will be described.
[0133]
In FIG. 19, the electron beam exiting the electron gun 31 passes through the magnetic lens 32 and the objective lens 33, and is narrowed down to a beam diameter of about the pixel size on the sample surface. At this time, a negative potential is applied to the sample by the ground electrode 37 and the retarding electrode 38, and the electron beam is decelerated between the objective lens and the sample, thereby achieving high resolution in a low acceleration voltage region. When the electron beam is irradiated, electrons are generated from the sample (wafer 1). By detecting the electrons generated from the sample in synchronization with the scanning of the electron beam in the X direction by the deflector 34 and the continuous movement of the sample in the Y direction by the stage 2, a two-dimensional electron beam image of the sample is obtained. can get. Electrons generated from the sample are captured by the detector 35 and amplified by the amplifier 36. Here, in order to enable high-speed inspection, an electrostatic deflector having a high deflection speed is used as the deflector 34 that repeatedly scans the electron beam in the X direction, and the electron gun 31 is an electron beam. It is desirable to use a thermal field emission electron gun that can shorten the irradiation time because the current can be increased, and it is desirable to use a semiconductor detector capable of high-speed driving as the detector 35.
[0134]
Next, the image extracting unit 102 will be described.
[0135]
The signal amplified by the amplifier 36 is converted into a digital signal by the AD converter 39, and then the preprocessing circuit 40 performs dark level correction (the average of the gradation values of a specific number of pixels during the beam blanking period is darkened). Level), correction of fluctuation of electron beam current (detection of beam current with an objective aperture not shown, normalization of signal with beam current), shading correction (correction of light intensity variation due to beam scanning position) Thereafter, the image quality is improved by performing a filtering process using a Gaussian filter, an average value filter, an edge enhancement filter, or the like. In addition, image distortion is corrected as necessary. These pre-processes are for converting the detected image in an advantageous manner in the subsequent defect determination process.
[0136]
The delay circuit 41 delays the signal by a predetermined time. If the delay time is set to a time for which, for example, the stage 2 moves by the chip pitch, the delayed signal g0 and the non-delayed signal f0 are the same in the adjacent chips. And the above-described chip comparison inspection. Alternatively, if the delay time is set to a time during which the stage 2 moves by the pitch of the memory cell, the delayed signal g0 and the undelayed signal f0 become image signals at the same location of the adjacent memory cells, and the cell comparison described above. It is an inspection. As described above, the image signals f0 and g0 to be compared are extracted from the image extracting unit 102. Hereinafter, f0 is referred to as a detected image, and g0 is referred to as a comparative image.
[0137]
Next, processing in the image processing unit 103 will be described.
[0138]
In the pixel-unit alignment unit 42, when the position of the detected image is used as a reference, the position where the “alignment” between the detected image and the comparative image becomes maximum is between 0 and 1 pixel. Shift the position.
[0139]
Next, the filters F and F ′ that minimize the brightness mismatch between images in the brightness matching filter operation unit are obtained. As mentioned above, solving the filter parameter coefficients using the least squares method, as shown in equations (7) and (8), dx0 and dy0 are obtained by obtaining various statistics ΣΣXX There is a need. The statistic calculator 44 calculates these various statistics, and calculates α and β by calculating (Equation 7) and (Equation 8) using ΣΣxx calculated by the sub CPU 45.
[0140]
The local gradation conversion circuit 46 is for performing gradation conversion that matches the brightness of f1 and g1 described above.
[0141]
The difference extraction circuit 49 obtains a difference image sub (x, y) between f1 and g1. This is expressed as follows.
[0142]
[Expression 21]

[0143]
The threshold value calculation circuit 48 uses the image signals f1, g1 and α, β that have passed through the delay circuits 46, 47 to determine whether the defect candidate is a defect candidate according to the value of sub (x, y). Threshold values thH (x, y) and thL (x, y) are calculated. thH (x, y) is a threshold value that defines the upper limit of sub (x, y), and thL (x, y) is a threshold value that defines the lower limit of sub (x, y). The configuration of the threshold value operation circuit 48 is shown in FIG. The contents of the calculation are expressed as follows.
[0144]
[Expression 22]

[0145]
[Expression 23]

[0146]
However,
[0147]
[Expression 24]

[0148]
[Expression 25]

[0149]
[Equation 26]

[0150]
Where a a and b b are real numbers from 0 to 0.5, γ is a real number greater than or equal to 0, and ε is an integer greater than or equal to 0
[0151]
[Expression 27]

[0152]
The first term A (x, y) on the right side of (Equation 22) and (Equation 23) for calculating the threshold value is used to correct the threshold value corresponding to α and β obtained by the positional deviation detection unit 43. It is a term of. For example, dx1 represented by (Equation 27) is regarded as a local change rate in the x direction of the gradation value of f1, and dx1 (x, y) * α is the gradation of f1 when the position is shifted by α. It can be referred to as a predicted value change. Therefore, the first term {dx1 (x, y) * α-dx2 (x, y) * (-α)} of A (x, y) indicates the position of f1 in the x direction as α and the position of g1 as − It can be referred to as a predicted value for each pixel how much the gradation value of the difference image between f1 and g1 changes when α is shifted. Similarly, the second term can be said to be a value predicted in the y direction. The first threshold value term A (x, y) is a term for canceling α and β.
[0153]
The second term B (x, y) on the right side of (Equation 22) and (Equation 23) for calculating the threshold value is used to allow a slight positional deviation of the pattern edge, a minute difference in pattern shape, and a pattern distortion. Term. As is clear from the comparison between (Equation 24) for obtaining A (x, y) and (Equation 25) for obtaining B (x, y), B (x, y) is the difference image of aa and bb. This is the absolute value of the change prediction of the gradation value. If A (x, y) cancels a known misregistration (considered), adding B (x, y) to A (x, y) This means that the position is shifted (considered) by aa in the x direction and bb in the y direction. That is, B (x, y) allows a positional shift of a a in the x direction and b b in the y direction.
[0154]
Further, subtracting B (x, y) from A (x, y) means that the position is further shifted by −a a in the x direction and −b b in the y direction from the aligned state. It is -B (x, y) that allows a positional shift of -a a in the x direction and -b b in the y direction. By providing an upper limit and a lower limit for the threshold value, positional deviations of ± a a and ± b b are allowed. Then, by setting the values of the parameters a a and b b to appropriate values, it is possible to freely control the allowable positional deviation amount.
[0155]
The third term C (x, y) in (Equation 22) and (Equation 23) for calculating the threshold value is a term for allowing a minute difference in gradation values. The addition of C (x, y) means that the gradation value of g1 is allowed to be larger by C (x, y) than the gradation value of f1, and the subtraction of C (x, y) is This means that the gradation value of g1 is allowed to be smaller by C (x, y) than the gradation value of f1. Here, C (x, y) is the sum of the representative value of the gradation value in the local region (here, the max value) multiplied by the proportionality constant γ and the constant ε, but stick to this function. It is not necessary, and if the method of changing the gradation value is known, it is better to use a function that matches it. For example, if it is known that the fluctuation range is proportional to the square root of the gradation value, C (x, y) = (max1 + max2) 1/2 * γ + ε should be used instead of (Equation 26). Similar to B (x, y), it is possible to freely control the difference in allowable gradation values by the parameters γ and ε.
[0156]
In the defect determination circuit 50, the output sub (x, y) of the difference extraction circuit 49 and the outputs thL (x, y) and thH (x, y) of the threshold value calculation circuit 48 are used.
[0157]
[Expression 28]

[0158]
Is satisfied, the pixel at the position (x, y) is determined as a non-defective candidate, and if not, the pixel at the position (x, y) is determined as a defective candidate. The defect determination circuit 50 outputs def (x, y) having a value of 0 for non-defective candidate pixels and 1 or more for defective candidate pixels.
[0159]
In the feature extraction circuit 50a, a defect that removes a noisy output by noise removal processing (for example, reduction / expansion processing is performed on def (x, y)), and then integrates defect candidate portions in the vicinity into one Perform candidate part merge processing. After that, feature quantities such as barycentric coordinates, XY projection length, and area are calculated for each group.
[0160]
The overall control unit 104 converts the coordinates of the defect portion into a coordinate system on the sample, deletes the pseudo defect, and finally collects defect data including the position on the sample and the feature amount.
[0161]
This defect data can be displayed on the display as in the first embodiment, or can be output by output means.
[0162]
Also, by displaying the image or image data before gradation conversion and the image or image data after gradation conversion on the display, respectively, or the image or image data after gradation conversion is displayed. By displaying, it is possible for the operator to visually confirm that the gradation conversion is correctly performed and the inspection is performed.
[0163]
According to the present embodiment, since a positional deviation as a whole small area, a minute positional deviation of individual pattern edges, and a minute difference in gradation values are allowed, a normal part is erroneously recognized as a defect. Disappears. In addition, it is possible to easily control the allowable amount of positional deviation and gradation value variation by the parameters a a, b b, γ, and ε.
[0164]
[Embodiment 3]
A third embodiment of the pattern inspection method and apparatus according to the present invention is shown in FIG. In FIG. 23, as in FIGS. 1 and 2, reference numeral 1 denotes an image sensor, which outputs a grayscale image signal corresponding to the brightness of reflected light from the semiconductor wafer 4 as the pattern to be inspected, that is, the grayscale. Reference numeral 2 denotes an A / D converter for converting a grayscale image signal obtained from the image sensor 1 into a digital image signal 9; 3 a delay memory for delaying the grayscale image signal; 4 a semiconductor wafer having a pattern to be inspected; An X, Y, Z, and θ direction (rotation) moving stage on which a semiconductor wafer 4 having an inspection pattern is placed, 6 an objective lens for the semiconductor wafer 4, and 7 an illumination of the semiconductor wafer 4 having an inspection pattern. An illumination light source 8 reflects the illumination light and irradiates the semiconductor wafer 4 through the objective lens 6 and transmits a reflected light from the semiconductor wafer 4. A digital image signal the image signal is converted by the A / D converter. In this way, the illumination light from the illumination light source 7 is reflected, and the semiconductor wafer 4 is configured to perform, for example, bright field illumination through the objective lens 6.
[0165]
Further, 3 may be a delay memory that stores and delays the image signal 9 for one or a plurality of repeated cell pitches, or stores and delays the image signal 9 for one or more repeated chip pitches. A delay memory may be used.
[0166]
11 is used to align the digital image signal 9 and the delayed digital image signal 10, and here, a positional shift amount in which the grayscale difference is minimized in units of pixels is detected by a normalized correlation, and this positional shift amount is calculated. Based on this, one image is shifted to align the two images. Note that the reason for normalization is to reduce the influence of the difference in brightness between images to be aligned.
[0167]
That is, the stored image g (x, y) is moved with respect to the detected image f (x, y), and the position where the correlation value is maximized is obtained by the following equation.
[0168]
[Expression 29]

[0169]
[30]

[0170]
[31]

[0171]
[Expression 32]

[0172]
[Expression 33]

[0173]
Here, the images are continuously detected by the image sensor, but the image is divided into line units described later, and alignment is performed in these units. In the above equation, the detected image has dimensions of X × Y pixels.
Although not shown in the drawing, the above-described normalized correlation for obtaining the positional deviation of the image does not have to be performed for all the images. For example, the image is divided into K in the longitudinal direction of the image sensor, and each divided small image is divided. Of the images (size of X / K × Y pixels), a small image with information may be used. For determining whether there is information, for example, each small image is differentiated, the presence or absence of an edge is detected, and a small image with many edges is selected. For example, when the image sensor is a linear image sensor with a multi-tap configuration capable of parallel output, each tap output image corresponds to a small image. This concept is based on the fact that images output in parallel have the same positional deviation.
[0174]
A gradation conversion unit 13 converts the gradations of the image signals having different brightnesses so that the brightnesses of the image signals match. Here, linear transformation is performed for each individual pixel using a gain and an offset, and the brightness is matched.
[0175]
[Expression 34]

[0176]
[Expression 35]

[0177]
[Expression 36]

[0178]
[Expression 37]

[0179]
Reference numeral 12 denotes a brightness conversion unit that converts image signals having different brightnesses so that the brightness of the image signals is matched. Here, filter operations are performed collectively on the entire image to match the brightness.
[0180]
Then, the obtained image signals are compared in the comparison unit 14 to detect a mismatch as a defect.
[0181]
Reference numeral 23 denotes an image input unit for inputting two images to be compared. From this image, a scatter diagram creation unit 24 obtains a scatter diagram. In the scatter diagram, the vertical axis and the horizontal axis indicate the brightness of two images. Reference numeral 25 denotes a display unit that displays the obtained scatter diagram. Reference numeral 15 denotes input means. For example, a threshold value for binarizing the absolute value of the difference image is input, and a line segment of the input threshold value is plotted on a scatter diagram. From this scatter diagram, it is easy to judge the validity of the input threshold value. Further, a threshold suitable for an image can be determined with reference to the displayed scatter diagram. An example of a scatter diagram is shown in FIG.
[0182]
When W (x, y, dx, dy) = 1, it is shown as follows.
[0183]
[Formula 38]

[0184]
[39]

[0185]
In a scatter diagram, a line segment can be applied to the plotted data group by the least square method or the like, and an error that is a deviation from the line segment can be obtained.
[0186]
If the straight line is Y = m · f (x, y) + n, the least square method (m, n) is calculated by the following equation), and linear approximation is performed.
[0187]
[Formula 40]

[0188]
[Expression 41]

[0189]
The error from the straight line is obtained by the following equation, for example.
[0190]
[Expression 42]

[0191]
[Expression 43]

[0192]
Based on this error, a threshold value can be obtained and the obtained threshold value can be plotted on a scatter diagram. For example, the threshold value is a value proportional to the square root of Ve. FIG. 24 shows a configuration example for realizing this.
[0193]
Reference numeral 26 denotes a statistic calculation unit for calculating the above-described line segment fitting and an error from the line segment. 27 is a threshold value calculation part which calculates a threshold value from this obtained statistic. Of course, there is no problem even if the threshold value can be input by the user.
[0194]
The images used in the scatter diagram are two images to be compared, for example, images after pixel unit alignment. At each stage of image processing, two images can be input to the image input unit 23.
[0195]
26 and 27 show an example in which two images are processed based on the method shown in FIG. The object is a line and space pattern detected in the lower right part of the image. The upper left is an area without a pattern. The histogram of the image during each process and the statistics of the difference image are also shown. As can be seen from the histogram, the brightness of the two images does not match at the first stage. First, a correlation value of the image is obtained by normalization correlation, and a position where the correlation value is high is obtained, thereby aligning in units of pixels. Next, local brightness correction, which is local gradation conversion, is performed on the two aligned images. Finally, filtering for matching the brightness of the two images is performed to further improve the matching degree of the brightness of the images.
[0196]
28 to 30 show scatter diagrams of images at each processing stage. At the stage of alignment in pixel units, the brightness of the two images does not match, and therefore, there is a variation from the straight line in the scatter diagram, not the diagonal 45 ° line. However, after the processing of local brightness correction and filtering, which is local gradation conversion according to the present invention, the scatter diagram is distributed near a straight line, which is effective in the sense that the brightness of the two images is made uniform. I know that there is. The slope and intercept are the slope and intercept of the line segment fitted to the scatter diagram data. The slope, which is a measure of the degree of coincidence between the two images, was initially 0.705, but became 0.986 after local brightness correction, which is local tone conversion, and then 0.991 after the filtering process. It can be seen that the degree of coincidence is improved.
[0197]
Furthermore, the value of Ve, which represents the degree of coincidence between the two images, was initially 40.02, but became 8.598 after the local brightness correction, which is local gradation conversion, and after the filtering process. It becomes 7.477, and it can be seen that the degree of coincidence of brightness is improved. In addition, the Ve value image is displayed as an image obtained by calculating the error Ve by linear approximation for each region of 7 × 7 pixels including the periphery of each pixel, for example, instead of the entire image of Ve. FIG. 28 to FIG. 30 show this. From this image, it can be seen where the brightness adjustment error is large.
[0198]
FIGS. 31 to 33 show scatter diagrams regarding the local contrast of an image. Here, for the contrast, the maximum value and the minimum value are detected around each pixel, for example, 3 × 3 pixels, and indicate the maximum value-minimum value. The local contrast is also considered to be meaningful because the scatter diagram is distributed near a straight line after the processing of local gradation conversion and filtering according to the present invention. The slope, intercept, etc. have the same meaning as in the previous figure. The Ve value image is obtained by linearly approximating a 7 × 7 pixel region including the periphery of each pixel in a local contrast image.
[0199]
34 to 37 show examples of scatter diagrams and threshold values. In FIG. 34, since the two images are different, the threshold value is set large so as not to erroneously detect this. FIG. 35 is a scatter diagram after brightness correction, which is local gradation conversion according to the present invention. Since the degree of coincidence of two images is high, the set threshold value is small. FIG. 36 is a scatter diagram after brightness adjustment, and the threshold value is further reduced. In FIG. 37, one of the images is subjected to linear gradation conversion in units of image instead of each pixel, and the threshold value has an offset on the scatter diagram.
[0200]
FIG. 38 shows an example in which the gradation conversion for each image is a piecewise linear conversion. In this example, it is divided into two sections.
[0201]
As described above, the scatter diagram and the threshold value can be used extensively, for example, as a standard for determining defect detection sensitivity or the like, or for checking the validity of the set threshold value.
[0202]
Such scatter diagram creation, display, threshold calculation using scatter diagram data, and the like can be performed by detecting an image and using this image before starting the examination. Furthermore, it will be apparent that a more sensitive inspection can be realized if scatter diagram creation, threshold setting, etc. are performed for each image in synchronization with image detection. The image detection may be performed at the stage where each processing is completed. As described above, the image processing is realized by pipeline processing, but it can be applied to other configurations.
[0203]
FIG. 41 shows another processing result. In FIG. 40, first, alignment by image matching (normalized correlation) is performed. Next, brightness correction is performed. In brightness correction, the above-described brightness filtering and local brightness correction are performed simultaneously. This further simplifies the configuration.
[0204]
In the above example, the brightness of the image is corrected by obtaining the coefficients m, n, etc. in real time (in synchronization with image detection). As a simplified version, a means for obtaining the coefficients m and n for the brightness of each image and correcting the brightness by referring to these coefficients may be considered. Each image may be two images to be compared. That is, the brightness of two images to be compared is input, and the coefficients m and n are obtained. Then, brightness correction is performed based on the created reference table.
[0205]
Next, the size of the image will be described. The size of the image, that is, the unit of image alignment (matching) can be determined by the following method. First, the amount of positional deviation between the two images to be compared is obtained in a finely divided unit. This is shown in FIG. Detection is performed separately in the X and Y directions. When this position shift data is subjected to spectrum analysis, a waveform as shown in FIG. 40 is obtained. In the spectrum analysis diagram, the vertical axis represents the spectral density and the horizontal axis represents the frequency. In this figure, attention is paid to the highest frequency and the high density frequency. In this case, it is 0.011. This frequency is determined by, for example, apparatus characteristics such as stage running characteristics and vibration characteristics.
[0206]
The spectrum analysis result shows that the positional deviation of the two images repeats at this frequency. If 88 lines, which is the reciprocal number, are used as the unit of the image, that is, the unit of matching, the peak of the positional deviation is included in the image. When the to peak appears and the amount of positional deviation is large, it is difficult to perform highly accurate alignment. However, when the image unit is 1/4 of the reciprocal of this frequency, the amount of positional deviation can be reduced to ½ or less of the peak positional deviation. Furthermore, when the reciprocal of the frequency is set to 1/8, the amount of positional deviation can be reduced to 1/4 or less of the peak positional deviation.
[0207]
Thus, the finer the unit of the image, the more accurate the image alignment should be, but the less information about the pattern that should be included in the image, the resulting image alignment accuracy. Does not go up. Therefore, from the spectrum analysis result, the size of the image is limited based on the required alignment accuracy. From the viewpoint of securing pattern information, the space information of the pattern is dependent on the pattern to be compared. The lower limit of the size of the image can be determined based on (information on the area where no pattern is formed). FIG. 25 shows this configuration. In the above description, attention is paid to the highest frequency, but it is also effective to pay attention to the amount of positional deviation and pay attention to the frequency corresponding to the highest frequency.
[0208]
Note that the above can be performed separately for the x and y components, or may be performed with attention paid only to the advancing direction of the stage as in the case of an accumulation type linear image sensor.
[0209]
According to the first, second, and third embodiments of the present invention, it is possible to detect a defect with high sensitivity without being affected by the difference in brightness of the pattern depending on the location. In addition, in a dark region such as the memory mat portion 21, even a pattern whose brightness varies greatly can be inspected with high sensitivity. In addition, highly accurate image matching can be realized without being affected by vibration characteristics of the apparatus. Therefore, a highly reliable inspection can be realized as compared with the conventional case.
[0210]
As described above, the embodiment of the present invention has been described mainly for the comparative inspection method using an optical microscope or a scanning electron microscope, but also when used for image detection obtained by other infrared rays or X-rays, It goes without saying that it is equally effective.
[0211]
【The invention's effect】
According to the configuration of the present invention, there is an effect that defects can be detected with high sensitivity without the overall inspection sensitivity being restricted by the difference in brightness caused by the difference in film thickness of the multilayer pattern.
[0212]
In addition, it is effective in realizing high-precision image matching. In addition, use scatter plots with the vertical and horizontal axes of brightness information obtained by performing processing such as brightness and differentiation of the two detected image signals to be compared, or brightness information during specified image processing. This makes it possible to determine the threshold value and validate it.
[Brief description of the drawings]
FIG. 1 is a configuration diagram of a defect inspection apparatus for a pattern to be inspected according to an embodiment of the present invention.
FIG. 2 is a configuration diagram of a defect inspection apparatus for a pattern to be inspected according to an embodiment of the present invention.
FIG. 3 is an explanatory diagram of an image brightness matching filter operation unit.
FIG. 4 is an explanatory diagram of a gradation conversion unit.
FIG. 5 is a diagram illustrating an example of a dual filter.
FIG. 6 is a diagram showing an example of a detected image and a difference image according to the present invention.
FIG. 7 is a diagram illustrating an example in which gradation conversion according to the present invention is performed.
FIG. 8 is a diagram illustrating an example in which gradation conversion according to the present invention is performed.
FIG. 9 is a diagram illustrating an example in which gradation conversion of the present invention is performed.
FIG. 10 is a diagram illustrating an example in which gradation conversion of the present invention is performed.
FIG. 11 is a diagram illustrating an example in which gradation conversion according to the present invention is performed.
FIG. 12 is an explanatory diagram of a threshold setting method according to the present invention.
FIG. 13 is an explanatory diagram of a threshold setting method according to the present invention.
FIG. 14 is an operation explanatory diagram of an image brightness matching filter operation unit according to the present invention.
FIG. 15 is a schematic explanatory diagram of a memory mat portion and a peripheral circuit portion in a memory chip of a pattern to be inspected.
16 is a histogram of brightness in a memory mat portion and a peripheral circuit portion in the memory chip of FIG.
FIG. 17 is a diagram illustrating a flow of CMP process processing.
FIG. 18 is a histogram of brightness in a memory mat portion and a peripheral circuit portion in different memory chips subjected to CMP processing;
FIG. 19 is a configuration diagram of a defect inspection apparatus for a pattern to be inspected according to an embodiment of the present invention.
FIG. 20 is a diagram illustrating a configuration of a threshold value operation circuit.
FIG. 21 is a diagram illustrating a sum of squares of a difference between two images.
FIG. 22 is a diagram illustrating a sum of absolute values of differences between two images.
FIG. 23 is a configuration diagram of a defect inspection apparatus for a pattern to be inspected according to an embodiment of the present invention.
FIG. 24 is a diagram for explaining scatter diagram creation and display according to an embodiment of the present invention.
FIG. 25 is a configuration diagram of a defect inspection apparatus for a pattern to be inspected according to an embodiment of the present invention.
FIG. 26 is a diagram illustrating a result at each image processing stage regarding two images to be compared;
FIG. 27 is a diagram illustrating a result at each image processing stage regarding two images to be compared;
FIG. 28 is a diagram showing a scatter diagram at each image processing stage for two images to be compared.
FIG. 29 is a diagram showing a scatter diagram at each image processing stage for two images to be compared;
FIG. 30 is a diagram illustrating a scatter diagram at each image processing stage for two images to be compared.
FIG. 31 is a diagram showing a scatter diagram of local contrast at each image processing stage for two images to be compared.
FIG. 32 is a diagram showing a scatter diagram of local contrast at each image processing stage for two images to be compared.
FIG. 33 is a diagram illustrating a scatter diagram of local contrast at each image processing stage for two images to be compared.
FIG. 34 is a diagram illustrating an example of a scatter diagram.
FIG. 35 is a diagram illustrating an example of a scatter diagram.
FIG. 36 is a diagram illustrating an example of a scatter diagram.
FIG. 37 is a diagram illustrating an example of a scatter diagram.
FIG. 38 is a diagram illustrating an example of a scatter diagram.
[Fig. 39] Fig. 39 is a diagram for describing the amount of image positional deviation.
FIG. 40 is a diagram for explaining spectrum analysis;
FIG. 41 is a diagram illustrating a result at each image processing stage regarding two images to be compared;
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... Image sensor, 2 ... A / D converter, 3 ... Delay memory, 4 ... Semiconductor wafer, 5 ... X, Y, Z, (theta) stage, 6 ... Objective lens, 7 ... Illumination light source, 8 ... Half mirror, 9 ... Image signal, 10 ... Image signal, 11 ... Pixel unit alignment unit, 12 ... Image brightness matching filter operation unit, 13 ... Tone conversion unit, 14 ... Comparison unit, 15 ... Input means, 16 ... CPU, 17 ... Memory 20: chip memory, 21: memory mat section, 22: peripheral circuit section (a non-repetitive pattern, including a repeated pattern having a plurality of repeated pitches), 23: image input section, 24: scatter diagram creating section , 25 ... display unit, 26 ... statistic calculation unit, 27 ... threshold value calculation unit (or input unit), 31 ... electron gun, 32 ... magnetic lens, 33 ... objective lens, 34 ... deflector, 35 ... detector 36 ... Amplifier 37 ... Land electrode, 38 ... retarding electrode, 39 ... A / D converter, 40 ... pre-processing circuit, 41 ... delay circuit, 42 ... pixel unit alignment unit, 46 ... local gradation conversion circuit, 48 ... threshold value Arithmetic circuit 49... Difference extraction circuit 50. Defect determination circuit 101. Detection unit 102 102 Image extraction unit 103 Image processing unit 104 Overall control unit

Claims (10)

A method for inspecting defects in a plurality of sets of patterns originally formed to be identical on a substrate, wherein a first image of the first pattern to be inspected is detected by detecting a first pattern to be inspected. And storing the first image, detecting a second pattern to be inspected to obtain a second image of the second pattern to be inspected, and storing the stored first image and the second image A first image obtained by performing the adjustment of the brightness of the entire image and the adjustment of the local brightness, and performing the adjustment of the brightness of the entire image and the adjustment of the local brightness. Inspecting the pattern by comparing with a second image,
To adjust the brightness of the entire image of the stored first image and the second image, a filter operation is performed collectively on the entire image of the stored first image and the second image. defect inspection method of an inspection pattern you and performing by. A method for inspecting defects in a plurality of sets of patterns originally formed to be identical on a substrate, wherein a first image of the first pattern to be inspected is detected by detecting a first pattern to be inspected. And storing the first image, detecting a second pattern to be inspected to obtain a second image of the second pattern to be inspected, and storing the stored first image and the second image A first image obtained by performing the adjustment of the brightness of the entire image and the adjustment of the local brightness, and performing the adjustment of the brightness of the entire image and the adjustment of the local brightness. Inspecting the pattern by comparing with a second image,
Matching the brightness of the entire image of the stored first image and the second image minimizes the square error of the density of the stored first image and the second image. The operation is performed using such a filter, and the brightness of the entire image is adjusted so that the square error in brightness between the first image and the second image is minimized. defect inspection method of an inspection pattern that. A method for inspecting defects in a plurality of sets of patterns originally formed to be identical on a substrate, wherein a first image of the first pattern to be inspected is detected by detecting a first pattern to be inspected. And storing the first image, detecting a second pattern to be inspected to obtain a second image of the second pattern to be inspected, and storing the stored first image and the second image A first image obtained by performing the adjustment of the brightness of the entire image and the adjustment of the local brightness, and performing the adjustment of the brightness of the entire image and the adjustment of the local brightness. Inspecting the pattern by comparing with a second image,
The brightness of the first image and the second image is obtained by performing linear transformation having a gain and an offset to locally match the brightness of the stored first image and the second image. Are converted to be almost the same,
A linear transformation with previous SL gain and offset, defect inspection of a pattern to be inspected you characterized be weighted in accordance with the brightness difference between the focused pixel and the surrounding pixels of the Komu locally combined area Method. The way to push locally align the brightness of the first image and the second image the storage, claims 1, characterized in that it is carried out in the previous alignment of pixels of two images 3 A defect inspection method for a pattern to be inspected according to any one of the above. That is intended to adjust the brightness of the first image and the second image the storage, no claim 1, characterized in that it is carried out in the following alignment of the pixel units of the two images to either of the 3 A method for inspecting a defect of a pattern to be inspected. A method for inspecting a defect of a plurality of sets of patterns originally formed to be identical on a substrate, the first image of the first pattern to be inspected by imaging a first pattern to be inspected And storing the first image, imaging the second pattern to be inspected to obtain a second image of the second pattern to be inspected, and storing the first image and the second image , The difference in local brightness between the first image and the second image in which the position shift is corrected is detected, and a local level is determined according to the detected difference in brightness. The local brightness is adjusted by performing tone conversion, and the brightness of the whole of the first image and the whole of the second image in which the local brightness is adjusted is adjusted. The pattern is inspected by comparing the first image and the second image combined with each other. ,
Matching the brightness of the whole of the first image and the whole of the second image to minimize the square error of the light and shade of the whole of the first image and the whole of the second image. the first total and the second defect inspection method of an inspection pattern you and performing by performing filtering on the whole and the image of the image to be. A method for inspecting a defect of a plurality of sets of patterns originally formed to be identical on a substrate, the first image of the first pattern to be inspected by imaging a first pattern to be inspected And storing the first image, imaging the second pattern to be inspected to obtain a second image of the second pattern to be inspected, and storing the first image and the second image , The difference in local brightness between the first image and the second image in which the position shift is corrected is detected, and a local level is determined according to the detected difference in brightness. The local brightness is adjusted by performing tone conversion, and the brightness of the whole of the first image and the whole of the second image in which the local brightness is adjusted is adjusted. The pattern is inspected by comparing the first image and the second image combined with each other. ,
The brightness of the first image and the second image is approximately adjusted by linear transformation having a gain and an offset to locally match the brightness of the first image and the second image. defect inspection method of an inspection pattern you and converting to be identical.   A defect inspection apparatus for inspecting defects of a plurality of sets of patterns originally formed to be identical on a substrate, imaging a first pattern to be inspected, and first of the first pattern to be inspected. An imaging means for obtaining a second image of the second pattern to be inspected by obtaining a second image to be inspected, and storing the first image obtained by imaging by the imaging means Storage means, gradation conversion means for locally converting at least one of the first image and the second image, and at least one of the first image and the second image is locally converted by the gradation conversion means. A brightness correction unit that matches the brightness of the first image and the second image that have undergone the tone conversion, and the first image and the second image that have been adjusted by the brightness correction unit are compared. And a defect detecting means for detecting a defect of the pattern. Defect inspection apparatus of an inspection pattern to symptoms. 9. The defect in a pattern to be inspected according to claim 8 , wherein the gradation converting means converts the brightness of the first image and the second image to be compared by linear conversion having a gain and an offset. Inspection device. Further, the image processing apparatus further includes misregistration correction means for correcting misregistration between the first image and the second image, and the gradation converting means corrects the misregistration by the misregistration correction means. 9. The defect inspection apparatus for a pattern to be inspected according to claim 8 , wherein gradation conversion is locally performed on at least one of the image and the image.

Images