JP3657565B2 - Image pattern identification and recognition method - Google Patents

Description

【００２２】
認識の判定結果としては、最大の類似度をもつ文字カテゴリｊを第１位に認識した文字カテゴリとして出力する。
この場合、類似度ＳH（Ｘ，Ｍjk）は文字線の方向や接続関係などの文字線構造に関する特徴量を抽出しなくてすむため、特徴抽出の処理が簡易になり、しかも、かすれ、つぶれ、文字背景の雑音などによって、これらの文字線構造が変形しても認識性能に影響を受けにくい特長を持っている。
なお、条件（３）の場合の認識処理では、上記類似度ＳH（Ｘ，Ｍjk）の絶対値｜ＳH（Ｘ，Ｍjk）｜を識別関数に用いてもよい。
【００２３】
次に、同定・認識・棄却判定処理の具体的な処理を図３を参照して説明する。図３は、ｎ＝１６（画素）、２文字カテゴリ（ｊ＝１，２）、Ｋ1＝１，Ｋ2＝２の場合の正規化文字パターンＸと２値マスクパターンＭjkとの例を示したものである。１１は２値マスクパターンＭ11（第１文字カテゴリの第１番目の２値マスクパターン）で、ｍｉ＝“１”となる個数Ｂ11＝８である。１２は２値マスクパターンＭ21（第２文字カテゴリの第１番目の２値マスクパターン）で、ｍi＝“１”となる個数Ｂ21＝７である。
１３は２値マスクパターンＭ22（第２文字カテゴリの第２番目の２値マスクパターン）で、ｍi＝“１”となる個数Ｂ22＝８である。２１〜２６は正規化文字パターンＸの６つのパターン例である。
なお、正規化文字パターン２５は正規化文字パターン２４の正規化文字パターンを白黒反転した白抜き文字パターンを示したものである。
【００２４】
初めに、正規化文字パターン２１の正規化文字パターンＸと２値マスクパターン１１〜１３の各２値マスクパターンＭjkとを照合する場合を説明する。ｘi＝“１”となる個数Ａ＝５で、ｘi＝“１”かつｍi＝“１”となる個数はそれぞれＣ11＝５，Ｃ21＝４，Ｃ22＝３となり、２値マスクパターン１１の２値マスクパターンＭ11だけが、条件（１）を満たすため、正規化文字パターン２１の正規化文字パターンＸは、第１文字カテゴリに同定される。
【００２５】
次に、正規化文字パターン２２の正規化文字パターンＸの場合は、Ａ＝６で、ｘi＝“１”かつｍi＝“１”となる個数はそれぞれＣ11＝５，Ｃ21＝６，Ｃ22＝５となり、２値マスクパターンＭ21だけが、条件（１）を満たすため、正規化文字パターン２２の正規化文字パターンＸは、第２文字カテゴリに同定される。
【００２６】
次に、正規化文字パターン２３の正規化文字パターンＸの場合は、Ａ＝４で、ｘi＝“１”かつｍi＝“１”となる個数はそれぞれＣ11＝４，Ｃ21＝４，Ｃ22＝３となり、２値マスクパターン１１の２値マスクパターンＭ11と２値マスクパターン１２の２値マスクパターンＭ21が、条件（２）を満たすため、正規化文字パターン２３の正規化文字パターンＸは、棄却される。
【００２７】
また、正規化文字パターン２４の正規化文字パターンＸの場合は、Ａ＝７で、ｘi＝“１”かつｍi＝“１”となる個数はそれぞれＣ11＝５，Ｃ21＝６，Ｃ22＝６となり、２値マスクパターン１１〜１３のいずれの２値マスクパターンも、条件（１）および（２）を満たさないため、条件（３）の識別関数が実行される。
【００２８】
この場合、正規化文字パターン２４の正規化文字パターンＸと２値マスクパターン１１〜１３の各２値マスクパターンＭjkとの類似度ＳHは、
【００２９】
【数２】

【００３０】
となり、２値マスクパターン１２の２値マスクパターンＭ21が最も類似度が高くなり、第２文字カテゴリが第１位の認識結果となる。
【００３１】
最後に、正規化パターン２５の正規化パターンＸ、すなわち正規化文字パターン２４を白黒反転した白ぬき文字パターンの場合は、Ａ＝９で、ｘi＝“１”かつｍi＝“１”となる個数は、それぞれＣ11＝３，Ｃ21＝１，Ｃ22＝２となり、２値マスクパターン１１〜１３のいずれの２値マスクパターンＭjkも条件（１）および（２）を満たさないため、条件（３）の識別関数が実行される。
ここで、識別関数として、上記の類似度ＳH（Ｘ，Ｍjk）に代えて、その絶対値｜ＳH（Ｘ，Ｍjk）｜を用いた場合、正規化文字パターン２５の正規化文字パターンＸと２値マスクパターン１１〜１３の各２値マスクパターンＭjkとの類似度｜ＳH｜は、
【００３２】
【数３】

【００３３】
となり、２値マスクパターン１２の２値マスクパターンＭ21が最も類似度が高くなり、正規化文字パターン２４の白黒反転パターンである正規化文字パターン２５に対しても、同様に第２文字カテゴリが第１位の認識結果となる。
【００３４】
次に、上述した条件（３）の認識処理（ステップ１０３）の他の実施例を説明する。次の実施例では、条件（３）の場合の認識処理における識別関数として、下式に示す本発明による他の新規な類似度Ｓc（Ｘ，Ｍjk）を用いる。
【００３５】
【数４】

【００３６】
この新規の類似度Ｓc（Ｘ，Ｍjk）の分子の項は、入力と２値マスクパターンで、 ｘi＝“１”かつｍi＝“１”となる個数Ｃjkとｘi＝“０”かつｍi＝“０”となる個数（ｎ−Ｂjk−Ｄjk）との積から、ｘi＝“０”かつｍi＝“１”となる個数（Ｂjk−Ｃjk）とｘi＝“１”かつｍi＝“０”となる個数Ｄjkとの積を差し引いた値から、
Ｃjk（ｎ−Ｂjk−Ｄjk）−（Ｂjk−Ｃjk）・Ｄjk＝ｎ・Ｃjk−Ａ・Ｂjk
と導いたものである。
図３に示す正規化文字パターン２４の正規化文字パターンＸと各２値マスクパターン１１〜１３の各２値マスクパターンＭjkとの類似度Ｓcは、
【００３７】
【数５】

【００３８】
となり、類似度ＳH（Ｘ，Ｍjk）を用いた場合と同様に、２値マスクパターン１２の２値マスクパターンＭ21が最も類似度が高くなり、第２文字カテゴリが第１位の認識結果となる。
【００３９】
また、正規化文字パターン２４を白黒反転した白抜き文字パターンである正規化文字パターン２５の正規化文字パターンＸの場合に、条件（３）の識別関数の実行において、識別関数として類似度Ｓc（Ｘ，Ｍjk）の絶対値｜Ｓc（Ｘ，Ｍjk）｜を用いたとき、正規化文字パターン２５と各２値マスクパターン１１〜１３との類似度｜Ｓc｜は、
【００４０】
【数６】

【００４１】
となり、２値マスクパターン１２の２値マスクパターンＭ21が最も類似度が高くなり、正規化文字パターン２４の場合と同様に、正規化文字パターン２５に対しても、第２文字カテゴリが第１位の認識結果となる。
【００４２】
次に、同定・棄却判定処理の他の方法を説明する。この方法では、上記の条件（１）〜（３）の判定において、識別関数に前述した新規の類似度Ｓc（Ｘ，Ｍjk）を用い処理を行う。
なお、以下の説明では、上記の条件（１）〜（３）に対応する条件を、それぞれ条件（１ａ）〜（３ａ）とする。
【００４３】
この同定・認識・棄却判定処理では、正規化文字パターンＸの要素ｘiと２値マスクパターンＭjkの要素ｍiとの間で、前記のＡ、Ｂjk、Ｃjkを計数し（ステップ１０１参照）、類似度Ｓc（Ｘ，Ｍjk）を計算し、予め２値マスクパターン記憶回路３に格納された各２値マスクパターンの同定しきい値ｔjkを用いて、以下に示す条件（１ａ）〜（３ａ）の処理を順次行う（ステップ１０２参照）。
なお、２値マスクパターンの同定しきい値ｔjkの設定方法の詳細については、後述する。
【００４４】
条件（１ａ）：
Ｓc（Ｘ，Ｍjk）≧ｔjkを満たす文字カテゴリｊがただ１個のカテゴリしか存在しない場合、入力文字は文字カテゴリｊに同定できたと判定する（ステップ１１０参照）。
【００４５】
条件（２ａ）：
Ｓc（Ｘ，Ｍjk）≧ｔjkを満たす文字カテゴリｊが２個以上存在する場合、入力文字は棄却と判定する（ステップ１２０参照）。
【００４６】
条件（３ａ）：
条件（１ａ）または条件（２ａ）を満す文字カテゴリが１個もない場合は、正規化文字パターンＸと２値マスクパターンＭjkとの間で、類似度Ｓcを識別関数として用いて認識処理を行い（ステップ１０３参照）、類似度Ｓc（Ｘ，Ｍjk）≧ｕjkを満たす文字カテゴリｊが存在する場合は、最も類似度が高い、２値マスクパターンＭjkの文字カテゴリを第１位の認識結果として出力する（ステップ１３０参照）。
他方、類似度Ｓc（Ｘ，Ｍjk）≧ｕjkを満たす文字カテゴリｊが存在しない場合、すなわち、すべての２値マスクパターンＭjkについて類似度Ｓc（Ｘ，Ｍjk）＜ｕjkとなる場合は、認識棄却と判定する（ステップ１２０参照）。
なお、変数ｕjkは、各２値マスクパターンＭjkについて、予め設定してある棄却判定のためのしきい値であり、同定しきい値ｔjkよりもさらに小さい値（ｕjk＜ｔjk）に設定されている。
【００４７】
ここで、同定、棄却判定処理の具体的な処理を再び図３を参照して説明する。初めに、正規化文字パターン２２の正規化文字パターンＸと２値マスクパターン１１〜１３の各２値マスクパターンＭjkと照合する場合を説明する。正規化文字パターン２２と各２値マスクパターン１１〜１３との類似度Ｓcは
【００４８】
【数７】

【００４９】
となり、２値マスクパターンＭ21だけが条件（１ａ）を満たすため、正規化文字パターン２２の正規化文字パターンＸは第２文字カテゴリに同定される。
次に、正規化文字パターン２６の場合は、２値マスクパターン１１〜１３の各２値マスクパターンとの類似度Ｓcは、
【００５０】
【数８】

【００５１】
となり、２値マスクパターン１１の２値マスクパターンＭ11と２値マスクパターン１２の２値マスクパターンＭ21が条件（２ａ）を満たすため、２６の正規化文字パターンＸは棄却される。
また、正規化文字パターン２３の場合は、各２値マスクパターン１１〜１３との類似度Ｓcは、
【００５２】
【数９】

【００５３】
となり、２値マスクパターン１１〜１３のいずれの２値マスクパターンも条件（１ａ）及び条件（２ａ）を満たさないため、条件（３ａ）の類似度Ｓcの識別関数が実行される。条件（３ａ）では、２値マスクパターン１２の２値マスクパターンＭ21が最も類似度が高いので、第２文字カテゴリが第１位の認識結果となる。
【００５４】
次に、誤認識した文字や未学習の文字等から２値マスクパターンを新規に作成し、２値マスクパターン記憶回路３の内容を追加・更新する処理について、二、三の実施例を説明する。
【００５５】
＜実施例１＞
これは、入力画像パターンの黒部または白部の情報をもとに、公知の識別関数として次式に示す類似度Ｓを用いて、反復学習により２値マスクパターンを作成する例である。
【００５６】
【数１０】

【００５７】
まず、画像入出力装置６から、追加・更新または新規に作成（ここでは追加とする）すべき文字カテゴリ（第ｊ文字カテゴリとする）に関する情報が２値マスクパターン作成回路５に入力される。２値マスクパターン作成回路５は、追加すべき入力画像に対して、正規化処理回路２から得られる、学習すべき第ｊ文字カテゴリに属するＮ個の入力画像パターン（正規化文字パターン）Ｘ1，Ｘ2，…，ＸNについて、以下のステップで新しい２値マスクパターンＭj(k+1)を作成する。
【００５８】
（ステップ１）
Ｎ個の入力画像パターンＸ1，Ｘ2，…，ＸNについて、ｘi＝“１”となる個数Ａの合計ＡTと画素位置ｉでｘi＝“１”となる入力画像パターンＸの合計個数Ａiを計数する。 Ａiを大きい順に並べ替え、その列をＡ’1，Ａ’2，…，Ａ’l，…，Ａ’nとおくと、予め設定したしきい値ａ（０＜ａ＜１）に対して
【００５９】
【数１１】

【００６０】
を満すＬを求める。ただし、Ｎが小さい場合、たとえばＮ＝１の場合はａ＝１とする。
【００６１】
（ステップ２）
ｌ＝１，２，…，Ｌの各順位に対応する画像位置ｉについて、２値マスクパターンＭj(k+1)の要素をｍi＝“１”とし、それ以外の画像位置ｉをｍi＝“０”に設定する。
【００６２】
（ステップ３）
Ｎ個の入力画像パターンＸ1，Ｘ2，…，ＸNの各画像Ｘtとステップ２で作成した２値マスクパターンＭj(k+1)との間で類似度Ｓ（Ｘt，Ｍj(k+1)）を計算し、最大類似度となる画像Ｘtを求め、その画像Ｘtを新しい２値マスクパターンＭj(k+1)に置き換える。
【００６３】
（ステップ４）
再度、Ｎ個の入力画像Ｘ1，Ｘ2，…，ＸNについて、各画像Ｘlと、すでに２値マスクパターン記憶回路３に記憶されているすべての２値マスクパターンＭjkおよびステップ３で作成した２値マスクパターンＭj(k+1)との間で類似度Ｓを計算し、認識する。
もし、Ｎ個の画像すべてが第ｊ文字カテゴリに認識されればステップ６へ、さもなければステップ５へいく。
【００６４】
（ステップ５）
第ｊ文字カテゴリに正しく認識されたＭ（＜Ｎ）個の画像について、ステップ１およびステップ２と同様の手順で２値マスクパターンＭj(k+1)を作成し、２値マスクパターン記憶回路３に記憶する。
残りの（Ｎ−Ｍ）個の画像については、ｋ＝ｋ＋１、Ｎ＝Ｎ−Ｍとしてステップ１へもどる。
【００６５】
（ステップ６）
ステップ１およびステップ２と同様の手順で２値マスクパターンＭj(k+1)を作成し、２値マスクパターン記憶回路３に記憶して終了とする。
【００６６】
＜実施例２＞
これは、入力画像パターン（サンプルパターン）の黒部または白部の情報をもとに、類似度ＳHを用いて、反復学習により新規に各文字カテゴリの２値マスクパターンおよび同定しきい値を作成する例である。
【００６７】
まず、画像入出力装置６から、新規に作成すべき文字カテゴリ（第ｊ文字カテゴリとする）に関する情報が２値マスクパターン作成回路５に入力される。２値マスクパターン作成回路５は入力画像に対して、正規化処理回路２から得られる、新規に作成すべき文字カテゴリｊ＝１，２，…，Ｊの各文字カテゴリについて、以下のステップでＫj個の２値マスクパターンＭj1，Ｍj2，…，Ｍjk，…，Ｍjkjを作成する。
【００６８】
（ステップ１ａ）
新規に作成すべき第ｊ文字カテゴリの２値マスクパターンの個数をＫj＝１とする。
第ｊ文字カテゴリに属するＮj個の入力画像パターン（正規化文字パターン）Ｘ1，Ｘ2，…ＸNjについて、ｘi＝“１”となる個数ＡTと画素位置ｉでｘi＝“１”となる個数Ａiを計数する。そして、画素位置ｉのｘi＝“１”となる個数の割合Ａi／Ｎjと全画素のｘi＝“１”となる個数の割合ＡT／（ｎＮj）を次式のように比べて、２値マスクパターンＭj1を作成する。
もし、Ａi／Ｎj≧ＡT／（ｎＮj）ならば、
Ｍj1の要素をｍi＝“１”に設定
さもなければ、
Ｍj1の要素をｍi＝“０”に設定
また、入力画像パターン（正規化文字パターン）Ｘ1，Ｘ2，…，ＸNjに対して２値マスクパターンＭj1の作成に用いたパターンとしてラベルｊ１を割り当てる。
新規に作成すべき文字カテゴリすべてについて、上記処理を行う。
【００６９】
（ステップ２ａ）
新規に作成すべき全入力画像パターンＸtについて、全文字カテゴリの全２値マスクパターンＭjkとの間で類似度ＳH（Ｘt，Ｍjk）を計算し、最大類似度をもつＭlhを求める。いま、すでに入力画像パターンＸtにラベルｊｋが割り当てられているとする。
もし、ｌ＝ｊならば、Ｘtは第ｊ文字カテゴリに正しく認識できたと判定して、新たにラベルｊｈを割り当てる。さもなければ、Ｘtは誤認識したと判定して、新たにラベルｊｅを割り当てる。
この処理終了後に、もし、全入力画像パターンＸtについて、誤認識したと判定するＸtがなければ、ステップ４ａへ進む。さもなければ、ステップ３ａへ進む。
【００７０】
（ステップ３ａ）
新規に作成すべき第ｊ文字カテゴリについてｈ＝１とおく。
ラベルｊｋが割り当てられているすべての入力画像パターンについて、もし、入力画像パターンが１つ以上存在すれば、ステップ１ａと同様の手順で新しい２値マスクパターンＭjhを作成し、ｈ＝ｈ＋１とおく。
第ｊ文字カテゴリのＫj個のラベルすべてについて、この判定処理を行う。
次に、もし、ラベルｊｅを割り当てられている入力画像パターンが１つ以上存在するならば、それらの入力画像パターンＸtすべてについて、ステップ１ａと同様の手順で、一時的な２値マスクパターンＭjeを作成し、すべてのＸtとＭjeとの間で類似度ＳH（Ｘt，Ｍje）を計算し、最大類似度をもつ入力画像パターンＸSを新規の２値マスクパターンＭjhとし、ｈ＝ｈ＋１とおく。
第ｊ文字カテゴリについて新しい２値マスクパターンの個数をＫj＝ｈとおく。
新規に作成すべきすべての文字カテゴリについて、この処理を繰り返し、新しい２値マスクパターンを２値マスクパターン記憶回路３に記憶する。そして、ステップ２ａへ戻る。
【００７１】
（ステップ４ａ）
各２値マスクパターンＭjkと第ｊ文字カテゴリ以外の全入力画像パターンＸtとの間で、類似度ＳH（Ｘt，Ｍjk）を計算し、各２値マスクパターンＭjkについての最大類似度ｔjkを求め、この最大類似度を２値マスクパターンＭjkの同定しきい値ｔjkに割り当て、これを２値マスクパターン記憶回路３に記憶する。全２値マスクパターンについて、これらの処理が終われば終了する。
【００７２】
＜実施例３＞
これは、類似度ＳHを用いることは実施例２と同様であるが、誤認識した文字や未学習文字の画像パターンを用いて２値マスクパターンおよび同定しきい値を更新する例である。
まず、画像入出力装置６から、更新すべき文字カテゴリ（第ｊ文字カテゴリとする）に関する情報が２値マスクパターン作成回路５に入力される。２値マスクパターン作成回路５は入力画像に対して、正規化処理回路２から得られる、更新すべき文字カテゴリｊ＝１，２，…，Ｊの各文字カテゴリについて、以下のステップで２値マスクパターンを更新する。
【００７３】
（ステップ１ｂ）
更新前の第ｊ文字カテゴリの２値マスクパターンの個数をＫjとする。
更新すべき全入力画像パターンＸtについて、全文字カテゴリの全２値マスクパターンＭjk（ｊ＝１，２…，Ｊ）（ｋ＝１，２…，Ｋj）との間で類似度ＳH（Ｘt，Ｍjk）を計算し、最大類似度をもつＭlhを求める。いま、入力画像パターンＸtが第ｊ文字カテゴリに属するとする。
もし、ｌ≠ｊならば、Ｘtは誤認識したと判定して、新たにラベルｊｅを割り当てる。
更新すべき全文字カテゴリについて、この処理終了後に、もし、全入力画像パターンＸtについて、誤認識したと判定するＸtがなければ、ステップ３ｂへ進む。さもなければ、ステップ２ｂへ進む。
【００７４】
（ステップ２ｂ）
ラベルｊｅが割り当てられた入力画像パターンＸtすべてについて、実施例２のステップ１ａと同様の手順で、一時的な２値マスクパターンＭjeを作成し、すべてのＸtとＭjeとの間で類似度ＳH（Ｘt，Ｍje）を計算し、最大類似度をもつ入力画像パターンＸsを追加すべき２値マスクパターンＭj(Kj+1)として２値マスクパターン記憶回路３に記憶する。また、Ｋj＝Ｋj＋１とおき、ステップ１ｂへ戻る。
【００７５】
（ステップ３ｂ）
各２値マスクパターンＭjkと第ｊ文字カテゴリ以外の全入力画像パターンＸtがあれば、そのＸtとの間で類似度ＳH（Ｘt，Ｍjk）を計算し、各２値マスクパターンＭjkについての最大類似度ｔ’jkを求め、この最大類似度がすべて２値マスクパターン記憶回路３に記憶されている同定しきい値ｔjk（新規に作成される場合はｔjk＝−ｎ）よりも大きい場合は、２値マスクパターンＭjkの同定しきい値ｔjk＝ｔ'jkに更新し、新しい値を２値マスクパターン記憶回路３に記憶する。全２値マスクパターンについて、これらの処理が終われば終了する。
【００７６】
ここで、第２および第３の実施例では、識別関数に類似度ＳHを用いた場合を述べたが、識別関数として、類似度ＳHの絶対値を用いることも可能である。
【００７７】
また、２値マスクパターンを作成する方法としては、これ以外にも公知の統計的推論による方法（たとえば「統計学」鈴木哲夫著（朝倉書店）や「Ｐattern classification and scence analysis」 Ｒ．Ｏ．Ｄuda and Ｐ．Ｅ．Ｈart著（Ｊohn Ｗiley &Ｓons Ｉns.））による方法を用いることも可能である。
また、第１の実施例においても、どの文字カテゴリにも含まれない画像パターンがあった場合にも、その画像パターンを用いて新たな文字カテゴリの２値マスクパターンを作成することが可能である。
【００７８】
さらに、以上の説明では文字を中心に示したが、図形などの一般の２値画像についても同様に処理可能である。
一例として、医療などで用いる多項目の診断結果をもとに、病気のカテゴリを認識する方法に適用する場合を、以下、簡単に説明する。
例えば、複数の診断項目の結果を、図３に示すような２値画像パターンとしてチェックシート上に表し（例えば、血圧の項目であれば、血圧が高ければ“１”、低ければ“０”としてその項目に該当する位置に“黒”または“白”で表す。）、このチェックシート上のパターンを２値画像パターンとして入力し、各病名のカテゴリの典型的な診断データに基づいて、予め記憶させておいた２値マスクパターンとの照合をとり、病名のカテゴリを認識する。このように、認識対象を２値画像化した後、上記の実施例のようにして認識することによって、病名のカテゴリの認識性能を向上させることができる。
【００７９】
【発明の効果】
以上説明したようにこの発明によれば、入力画像パターンと標準辞書に予め記憶してある２値マスクパターンとを比較して入力画像を同定・棄却または認識することにより、ＯＣＲの利用者が誤認識やリジェクトパターンを逐次、目視により再検査し修正していく作業において、同定された画像については再検査する必要が不要となり、仕事の負担が軽減される利点がある。
また、従来の文字認識法では、認識特徴の１要素が深さｎ（ｎ＞１）ビットで記憶する場合が多いが、本発明では２値マスクパターンの各要素は１ビットで表現できるので、記憶容量が１／ｎに削減できる。さらに、公知の画像符号化法を併用することで、２値マスクパターンの記憶容量を更に削減でき、処理効率が高く、辞書容量がコンパクトですむという利点を持つ。
【００８０】
また、本発明によれば、文字線の方向や接続関係などの文字線構造に関する特徴量を抽出しなくてすむため、特徴抽出の処理が簡易になり、しかも、かすれ、つぶれ、文字背景の雑音などによって、これらの文字構造が変形しても認識性能に影響を受けにくい利点がある。かすれ、つぶれ、文字背景の雑音だけでなく、文字部を黒点、背景を白点として構成される画像パターンから作成した２値マスクパターンを用いても、文字部が白点、背景が黒点として構成される、いわゆる白抜き文字についてのかすれ、つぶれ、文字背景の雑音などに対しても高い認識性能をもつ利点ある。
【００８１】
さらに、本発明によれば入力画像パターンと標準辞書に予め記憶してある２値マスクパターンとを比較して入力画像を同定・棄却または認識することにより、ＯＣＲの利用者が誤認識やリジェクトパターンを逐次、目視により再検査し修正していく作業において、認識された画像については再検査する必要が不要となり、仕事の負担が軽減される利点もある。
【図面の簡単な説明】
【図１】 本発明による画像パターンの同定・認識処理装置の構成例を示すブロック図である。
【図２】 本発明による同定・棄却・認識の判定処理の一例のフローチャートである。
【図３】 同定・棄却・認識の具体的判定処理を説明するためのパターン例を示す図である。
【符号の説明】
１ ２値画像パターン記憶回路
２ 正規化処理回路
３ ２値マスクパターン記憶回路
４ 認識判定回路
５ ２値マスクパターン作成回路
６ 画像入出力装置[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an image pattern identification / recognition method of a binary pattern in which an input pattern is represented by white or black. In the case of many character / graphic categories such as printed kanji, handwritten kanji, alphanumeric characters, symbols, mathematical formulas, and graphics, which are representative examples of binary patterns, images and stripes with noise such as blurring and crushing It can be applied to a character recognition method for recognizing a character image subjected to a design process consisting of, for example, a white character. It can also be applied to a method for recognizing multiple categories of failures and illnesses that occur in the system based on the diagnosis results of multiple items used in disasters and medical care.
[0002]
[Prior art]
Conventionally, character recognition processing devices and drawing recognition devices such as Kanji OCR (Optical Character Reader) extract features for recognition from characters and graphic patterns in the form of vectors, and each category in a standard dictionary created in advance. There is known a method in which an identification function such as a degree of similarity or a degree of difference is obtained with respect to a standard pattern vector, and the most similar character or figure category is used as a recognition result.
For feature vector elements such as conventional character recognition, features that reflect the character line structure such as the direction of the character line, the connection relationship, and the positional relationship are widely used, but the design is composed of unique textures. For an image with severe characters, blurring, blurring, character background noise, and the like, these feature amounts fluctuate greatly, and it is almost impossible to obtain sufficient recognition accuracy.
[0003]
In addition, a monochrome pattern input in conventional character recognition is used as an element of a feature vector, and a standard pattern vector (represented by a binary pattern similar to a feature vector) of each category in a standard dictionary created in advance is used together. Simple similarity is known as an identification function based on the number of black pixels. However, with this method, there is a problem that an image with a severe design character or character background noise composed of a unique texture is erroneously recognized as a category having a large number of black pixels in the standard pattern vector.
[0004]
[Problems to be solved by the invention]
As described above, conventional OCR and other recognition technologies have been well established for correctly recognizing characters that are misrecognized due to design characters, blurring, blurring, character background noise, white character mixing, etc. There was no problem.
An object of the present invention is to provide an image pattern identification / recognition method having a recognition function capable of dealing with design characters, severe character deformation such as blurring and crushing, and character strings mixed with white characters.
[0005]
[Means for Solving the Problems]
In order to achieve the above object, according to the first aspect of the present invention, an input binary image pattern is checked against each binary mask pattern for each recognition category that has been created in advance. An image pattern identification / recognition method for determining a recognition category to which a pattern belongs,
Black image of the binary image pattern X at a plurality of predetermined positions in a pattern composed of n pixels Elementary When counting the number A and the black pixel number Bjk of the binary mask pattern Mjk (j is the number of categories, k is the number of mask patterns), the pixels of the binary image pattern X value A first step of calculating a number Cjk (hereinafter referred to as a first degree of coincidence) in which both Xi and the pixel value mi of the binary mask pattern are black pixels;
A second step of calculating a number Djk (hereinafter referred to as a first degree of mismatch) in which the pixel value Xi of the binary image pattern X is a black pixel and the pixel value mi of the binary mask pattern is a white pixel;
By subtracting the first degree of coincidence calculated in the first step and the first degree of inconsistency calculated in the second step from the number of pixels n of the pattern, at the plurality of predetermined positions. A third step of calculating a number of white pixels in the binary image pattern and the binary mask pattern (hereinafter, second matching degree);
By subtracting the first degree of coincidence calculated in the first step from Bjk, the number that the binary image pattern becomes a white pixel and the binary mask pattern becomes a black pixel at the plurality of predetermined positions ( Hereinafter, the second inconsistency degree is calculated. First 4 steps,
A fifth step of calculating a third degree of coincidence by multiplying the first degree of coincidence and the second degree of coincidence;
A sixth step of calculating a third mismatch degree by multiplying the first mismatch degree and the second mismatch degree;
A first step configured by a seventh step of obtaining a similarity between the binary image pattern and the binary mask pattern by normalizing a value obtained by subtracting the third mismatch from the third match. Procedure and
A second procedure for recognizing that the binary image pattern belongs to a recognition category having the maximum similarity calculated by the first procedure;
It is characterized by having.
[0006]
According to a second aspect of the present invention, in the image pattern identification / recognition method according to the first aspect, all the positions in the pattern are used as a plurality of predetermined positions in the pattern.
[0007]
The invention according to claim 3 Term The image pattern identification / recognition method according to claim 1, further comprising a step of setting a threshold value for each binary mask pattern of each recognition category in advance of Change to the procedure
When the similarity in the first procedure between the binary image pattern and the binary mask pattern of the recognition category is equal to or higher than the threshold value of the binary mask pattern for only one recognition category. ,
The binary image pattern has a third procedure for identifying it in the recognition category.
[0008]
Further, the invention according to claim 4 is a claim. Term 3. The image pattern identification / recognition method according to 3, wherein the input binary image pattern is recognized by any of the third procedures. Knowledge The similarity in the first procedure between the binary image pattern and the binary mask pattern of the recognition category for two or more recognition categories does not belong to the category. A fourth procedure for rejecting the input binary image pattern when the threshold value is exceeded is provided.
[0009]
According to a fifth aspect of the present invention, there is provided the image pattern identification / recognition method according to any one of the first to fourth aspects, wherein the predetermined value is determined based on information on a black portion or a white portion of the input binary image pattern. A new binary mask pattern is created, added, or updated by iterative learning using the similarity of 5 It has the procedure of this.
[0010]
[Action]
According to the above configuration, the input binary image pattern is collated with each binary mask pattern created in advance for each recognition category, and the recognition category to which the binary image pattern belongs is efficiently identified, recognized or rejected. can do.
Also, a new binary mask pattern can be created, added, or updated by iterative learning based on information on the input image pattern.
[0011]
【Example】
An embodiment of the present invention will be described below with reference to the drawings.
FIG. 1 is a block diagram showing a configuration example of an image pattern identification / recognition processing apparatus to which the method of the present invention is applied. 1 is a binary image pattern storage circuit, 2 is a normalization processing circuit, and 3 is a binary mask pattern storage circuit. Reference numeral 4 denotes an identification / recognition determination circuit, 5 denotes a binary mask pattern creation circuit, and 6 denotes an image input / output device.
The binary image pattern storage circuit 1 stores a binary image pattern of characters or figures to be identified or recognized. The binary image pattern is composed of n pixels. For example, white pixels are “0” and black pixels are “1”.
[0012]
The normalization processing circuit 2 inputs a binary image pattern composed of n pixels extracted from the binary image pattern storage circuit 1, and uses, for example, a conventionally known center of gravity and second moment to determine the position and size. Perform normalization processing.
The binary mask pattern storage circuit 3 stores in advance one or a plurality (generally a plurality) of binary mask patterns for each category used for identification and recognition determination of the input image pattern.
[0013]
The identification / recognition determination circuit 4 inputs a binary image pattern of n pixels from the normalization processing circuit 2 and collates it with the binary mask pattern of each category of the binary mask pattern storage circuit 3, for example, an input image pattern Are included in the black portion (or white portion) of the binary mask pattern of a specific category, and the binary mask pattern black portion (or white portion) of the binary mask pattern of another category and the black portion (or white portion) of another category. If it is not included only in the overlapping part (or white part), it is identified and recognized as that particular category, and the black part (or white part) of the input image pattern is included only in the overlapping part In other cases, a recognition function such as a well-known similarity function or a recognition function that recognizes white text with high accuracy based on the white or black information of the input image pattern is used. Performs a rejection, and outputs the recognition result.
[0014]
The binary mask pattern creation circuit 5 receives the image information output from the image input / output device 6 and the normalized binary image pattern of the addition / update or new creation category from the normalization processing circuit 2 and receives the binary mask. Create a pattern. The binary mask pattern created by the binary mask pattern creation circuit 5 is newly registered in the binary mask pattern storage circuit 3 or added / updated to the contents of the binary mask pattern storage circuit 3.
[0015]
The image input / output device 6 inputs classification result information such as a character or figure category number, an identification function value, a binary image pattern, and the like output from the identification / recognition determination circuit 4 and displays the information. Input image information with a keyboard or mouse and display it. Further, the image input / output device 6 sends information necessary for newly creating or updating a binary mask pattern to the binary mask pattern creating circuit 5.
[0016]
First, input image pattern identification and recognition processing will be described with reference to FIG.
After the binary image pattern of the character to be identified / recognized is read from the binary image pattern storage circuit 1 and the character image pattern is normalized in the normalization processing circuit 2, the identification is performed. Input to the recognition determination circuit 4.
Here, the character image pattern normalized by the normalization processing circuit 2 is referred to as a normalized character pattern X. The normalized character pattern X can be expressed by a vector having n elements of X = (x1, x2,..., Xi,... Xn), and each element xi is an image position i (= 1, 2,..., N). (For example, xi = “0” is a white point and xi = “1” is a black point).
[0017]
The identification / recognition determination circuit 4 is the k-th (k = 1, 2,...) Of the jth character category (j = 1, 2,. .., Kj) are collated with the binary mask pattern Mjk. Mjk is represented by Mjk = (m1, m2,..., Mi,..., Mn) jk. Each mi takes a binary value. For example, if mi = “0”, it represents a white point, and if mi = “1”, it represents a black point.
[0018]
As an example of collation, a case where the normalized character pattern X and the binary mask pattern Mjk are collated will be described. FIG. 2 shows an example of a processing flowchart.
First, between the element xi of the normalized character pattern X and the element mi of the binary mask pattern Mjk,
(A) the number A of x i = “1”;
(B) the number Bjk where mi = “1”;
(C) the number Cjk where x i = “1” and mi = “1”;
(D) the number Djk where x i = “1” and mi = “0”;
Are counted (step 101).
[0019]
Next, determination processing is sequentially performed from the following condition (1) to condition (3) to determine identification, rejection, and recognition (step 102).
Condition (1):
If there is only one category of character category j satisfying Cjk ≦ Bjk and A = Cjk, it is determined that the input character can be identified as character category j (step 110).
Condition (2):
If there are two or more character categories j satisfying Cjk ≦ Bjk and A = Cjk, the input character is determined to be rejected (step 120).
Condition (3):
When there is no character category satisfying the conditions (1) and (2), a character with a strong background noise or a white character is recognized between the normalized character pattern X and the binary mask pattern Mjk. Recognition processing is performed using an identification function or a known identification function (step 103), and recognition or rejection is determined (step 104). If recognized, the character category is set as a candidate category (step 130).
[0020]
Here, as an example of the recognition process (step 103) in the case of the condition (3), a case where the new similarity SH (X, Mjk) according to the present invention is used for the discrimination function will be described. That is, the similarity SH (X, Mjk) expressed by the following equation is calculated for the black pixel number A of the normalized character pattern X.
[0021]
[Expression 1]

[0022]
As a recognition determination result, the character category j having the maximum similarity is output as the character category recognized first.
In this case, the similarity SH (X, Mjk) eliminates the need to extract the feature quantity related to the character line structure such as the direction of the character line and the connection relationship, so that the feature extraction process is simplified, and the blurring, squashing, Even if these character line structures are deformed due to noise of the character background, etc., it has the feature that it is hardly affected by the recognition performance.
In the recognition process in the case of the condition (3), the absolute value | SH (X, Mjk) | of the similarity SH (X, Mjk) may be used as a discrimination function.
[0023]
Next, specific processing of identification / recognition / rejection determination processing will be described with reference to FIG. FIG. 3 shows an example of a normalized character pattern X and a binary mask pattern Mjk when n = 16 (pixels), two character categories (j = 1, 2), K1 = 1, and K2 = 2. It is. 11 is a binary mask pattern M11 (the first binary mask pattern of the first character category), and the number B11 = 8 where mi = “1”. Reference numeral 12 denotes a binary mask pattern M21 (first binary mask pattern of the second character category), where the number B21 = 7 where mi = “1”.
Reference numeral 13 denotes a binary mask pattern M22 (the second binary mask pattern of the second character category), where the number B22 = 8 where mi = “1”. 21 to 26 are six pattern examples of the normalized character pattern X.
The normalized character pattern 25 indicates a white character pattern obtained by reversing the normalized character pattern of the normalized character pattern 24 in black and white.
[0024]
First, a case where the normalized character pattern X of the normalized character pattern 21 and the binary mask patterns Mjk of the binary mask patterns 11 to 13 are collated will be described. The number A = 5 where xi = “1”, the number where xi = “1” and mi = “1” is C11 = 5, C21 = 4, C22 = 3, respectively. Since only the mask pattern M11 satisfies the condition (1), the normalized character pattern X of the normalized character pattern 21 is identified in the first character category.
[0025]
Next, in the case of the normalized character pattern X of the normalized character pattern 22, the numbers of A = 6, xi = “1”, and mi = “1” are C11 = 5, C21 = 6, C22 = 5, respectively. Since only the binary mask pattern M21 satisfies the condition (1), the normalized character pattern X of the normalized character pattern 22 is identified in the second character category.
[0026]
Next, in the case of the normalized character pattern X of the normalized character pattern 23, the number of A = 4, xi = “1” and mi = “1” is C11 = 4, C21 = 4, C22 = 3, respectively. Since the binary mask pattern M11 of the binary mask pattern 11 and the binary mask pattern M21 of the binary mask pattern 12 satisfy the condition (2), the normalized character pattern X of the normalized character pattern 23 is rejected. The
[0027]
In the case of the normalized character pattern X of the normalized character pattern 24, A = 7, the number of xi = “1” and mi = “1” is C11 = 5, C21 = 6, and C22 = 6, respectively. Since any of the binary mask patterns 11 to 13 does not satisfy the conditions (1) and (2), the discrimination function of the condition (3) is executed.
[0028]
In this case, the similarity SH between the normalized character pattern X of the normalized character pattern 24 and each of the binary mask patterns Mjk of the binary mask patterns 11 to 13 is expressed as follows:
[0029]
[Expression 2]

[0030]
Thus, the binary mask pattern M21 of the binary mask pattern 12 has the highest similarity, and the second character category is the first recognition result.
[0031]
Finally, in the case of a normalization pattern X of the normalization pattern 25, that is, a white character pattern obtained by reversing the normalization character pattern 24 in black and white, the number of Ai = 9, xi = “1”, and mi = “1” Respectively, C11 = 3, C21 = 1, C22 = 2, and any of the binary mask patterns Mjk of the binary mask patterns 11 to 13 does not satisfy the conditions (1) and (2). An identification function is executed.
Here, when the absolute value | SH (X, Mjk) | is used in place of the similarity SH (X, Mjk) as the discrimination function, the normalized character patterns X and 2 of the normalized character pattern 25 are used. The similarity | SH | of each of the value mask patterns 11 to 13 with each of the binary mask patterns Mjk is
[0032]
[Equation 3]

[0033]
Thus, the binary mask pattern M21 of the binary mask pattern 12 has the highest similarity, and the second character category is the same for the normalized character pattern 25 which is the black and white inversion pattern of the normalized character pattern 24. The first recognition result.
[0034]
Next, another embodiment of the recognition process (step 103) for the above condition (3) will be described. In the next embodiment, another novel similarity score Sc (X, Mjk) according to the present invention shown in the following equation is used as the discrimination function in the recognition process in the case of the condition (3).
[0035]
[Expression 4]

[0036]
The numerator term of this new similarity score Sc (X, Mjk) is the number of input Cjk, xi = “1” and mi = “1”, xi = “0” and mi = “ From the product of the number (n−Bjk−Djk) which becomes 0 ”, the number (Bjk−Cjk) where xi =“ 0 ”and mi =“ 1 ”, xi =“ 1 ”and mi =“ 0 ”. From the value obtained by subtracting the product from the number Djk,
Cjk (n-Bjk-Djk)-(Bjk-Cjk) .Djk = n.Cjk-A.Bjk
It is what led.
The similarity Sc between the normalized character pattern X of the normalized character pattern 24 shown in FIG. 3 and each of the binary mask patterns Mjk of each of the binary mask patterns 11 to 13 is expressed as follows:
[0037]
[Equation 5]

[0038]
As in the case of using the similarity SH (X, Mjk), the binary mask pattern M21 of the binary mask pattern 12 has the highest similarity, and the second character category is the first recognition result. .
[0039]
Further, in the case of the normalized character pattern X of the normalized character pattern 25 which is a white character pattern obtained by reversing the normalized character pattern 24 in black and white, the similarity score Sc ( When the absolute value | Sc (X, Mjk) | of X, Mjk) is used, the similarity | Sc | between the normalized character pattern 25 and each of the binary mask patterns 11 to 13 is
[0040]
[Formula 6]

[0041]
Thus, the binary mask pattern M21 of the binary mask pattern 12 has the highest similarity, and the second character category is ranked first for the normalized character pattern 25 as in the case of the normalized character pattern 24. It becomes the recognition result.
[0042]
Next, another method of identification / rejection determination processing will be described. In this method, in the determination of the above conditions (1) to (3), processing is performed using the above-described new similarity score Sc (X, Mjk) as the discrimination function.
In the following description, conditions corresponding to the above conditions (1) to (3) are referred to as conditions (1a) to (3a), respectively.
[0043]
In this identification / recognition / rejection determination process, A, Bjk, and Cjk are counted between the element xi of the normalized character pattern X and the element mi of the binary mask pattern Mjk (see step 101), and the degree of similarity. Sc (X, Mjk) is calculated, and the following conditions (1a) to (3a) are processed using the identification threshold value tjk of each binary mask pattern stored in the binary mask pattern storage circuit 3 in advance. Are sequentially performed (see step 102).
Details of the method of setting the identification threshold value tjk of the binary mask pattern will be described later.
[0044]
Condition (1a):
If there is only one category of character category j satisfying Sc (X, Mjk) ≧ tjk, it is determined that the input character can be identified as character category j (see step 110).
[0045]
Condition (2a):
If there are two or more character categories j satisfying Sc (X, Mjk) ≧ tjk, the input character is determined to be rejected (see step 120).
[0046]
Condition (3a):
If no character category satisfies the condition (1a) or the condition (2a), the recognition process is performed between the normalized character pattern X and the binary mask pattern Mjk using the similarity score Sc as an identification function. If there is a character category j satisfying the similarity Sc (X, Mjk) ≧ ujk (see step 103), the character category of the binary mask pattern Mjk having the highest similarity is set as the first recognition result. Output (see step 130).
On the other hand, when there is no character category j satisfying the similarity Sc (X, Mjk) ≧ ujk, that is, when the similarity Sc (X, Mjk) <ujk is satisfied for all binary mask patterns Mjk, the recognition is rejected. Determine (see step 120).
The variable ujk is a preset threshold value for rejection determination for each binary mask pattern Mjk, and is set to a value (ujk <tjk) smaller than the identification threshold value tjk. .
[0047]
Here, the specific process of the identification and rejection determination process will be described with reference to FIG. 3 again. First, a case in which the normalized character pattern X of the normalized character pattern 22 is compared with the binary mask patterns Mjk of the binary mask patterns 11 to 13 will be described. The similarity Sc between the normalized character pattern 22 and each of the binary mask patterns 11 to 13 is
[0048]
[Expression 7]

[0049]
Since only the binary mask pattern M21 satisfies the condition (1a), the normalized character pattern X of the normalized character pattern 22 is identified as the second character category.
Next, in the case of the normalized character pattern 26, the similarity Sc with each of the binary mask patterns of the binary mask patterns 11 to 13 is
[0050]
[Equation 8]

[0051]
Since the binary mask pattern M11 of the binary mask pattern 11 and the binary mask pattern M21 of the binary mask pattern 12 satisfy the condition (2a), the 26 normalized character patterns X are rejected.
In the case of the normalized character pattern 23, the similarity Sc with each of the binary mask patterns 11 to 13 is
[0052]
[Equation 9]

[0053]
Thus, since none of the binary mask patterns 11 to 13 satisfies the conditions (1a) and (2a), the discrimination function of the similarity score Sc of the condition (3a) is executed. Under condition (3a), the binary mask pattern M21 of the binary mask pattern 12 has the highest similarity, so that the second character category is the first recognition result.
[0054]
Next, a couple of embodiments will be described for the process of newly creating a binary mask pattern from misrecognized characters, unlearned characters, etc. and adding / updating the contents of the binary mask pattern storage circuit 3. .
[0055]
<Example 1>
This is an example of creating a binary mask pattern by iterative learning using the similarity S shown in the following equation as a known discriminant function based on the information on the black part or white part of the input image pattern.
[0056]
[Expression 10]

[0057]
First, from the image input / output device 6, information relating to a character category (referred to as a jth character category) to be added / updated or newly created (here, added) is input to the binary mask pattern creating circuit 5. The binary mask pattern generation circuit 5 obtains N input image patterns (normalized character patterns) X1, which belong to the jth character category to be learned, obtained from the normalization processing circuit 2 for the input image to be added. For X2,..., XN, a new binary mask pattern Mj (k + 1) is created in the following steps.
[0058]
(Step 1)
For N input image patterns X1, X2,..., XN, the total number AT of the number A where xi = “1” and the total number Ai of the input image patterns X where xi = “1” at the pixel position i are counted. . When Ai is rearranged in the descending order and the columns are set as A′1, A′2,..., A′1, A′n, a predetermined threshold value a (0 <a <1) is set.
[0059]
[Expression 11]

[0060]
L that satisfies However, when N is small, for example, when N = 1, a = 1.
[0061]
(Step 2)
For the image position i corresponding to each rank of l = 1, 2,..., L, the element of the binary mask pattern Mj (k + 1) is set to mi = “1”, and the other image position i is set to mi = “ Set to 0 ”.
[0062]
(Step 3)
Similarity S (Xt, Mj (k + 1)) between each image Xt of N input image patterns X1, X2,..., XN and the binary mask pattern Mj (k + 1) created in step 2 Is calculated to obtain an image Xt having the maximum similarity, and the image Xt is replaced with a new binary mask pattern Mj (k + 1).
[0063]
(Step 4)
Again, for N input images X1, X2,..., XN, each image Xl, all binary mask patterns Mjk already stored in the binary mask pattern storage circuit 3, and the binary mask created in step 3 are used. The similarity S is calculated and recognized with respect to the pattern Mj (k + 1).
If all N images are recognized in the jth character category, go to step 6, otherwise go to step 5.
[0064]
(Step 5)
A binary mask pattern Mj (k + 1) is created for the M (<N) images correctly recognized in the j-th character category by the same procedure as Step 1 and Step 2, and the binary mask pattern storage circuit 3 To remember.
For the remaining (N−M) images, go back to step 1 with k = k + 1 and N = N−M.
[0065]
(Step 6)
A binary mask pattern Mj (k + 1) is created in the same procedure as in step 1 and step 2, stored in the binary mask pattern storage circuit 3, and the process ends.
[0066]
<Example 2>
This creates a new binary mask pattern and identification threshold value for each character category by iterative learning using the similarity SH based on the black or white information of the input image pattern (sample pattern). It is an example.
[0067]
First, information about a character category to be newly created (referred to as a jth character category) is input from the image input / output device 6 to the binary mask pattern creation circuit 5. The binary mask pattern creation circuit 5 performs the following steps for each character category j = 1, 2,..., J to be newly created, which is obtained from the normalization processing circuit 2 for the input image. , Mjk,..., Mjkj are created.
[0068]
(Step 1a)
Let Kj = 1 be the number of binary mask patterns of the jth character category to be newly created.
For Nj input image patterns (normalized character patterns) X1, X2,... XNj belonging to the jth character category, the number AT where xi = “1” and the number Ai where xi = “1” at the pixel position i are set. Count. Then, the ratio Ai / Nj of the number xi = “1” of the pixel position i and the ratio AT / (nNj) of the number xi = “1” of all the pixels are compared as in the following equation, and the binary mask A pattern Mj1 is created.
If Ai / Nj ≧ AT / (nNj),
Set the element of Mj1 to mi = "1"
Otherwise,
Set the element of Mj1 to mi = “0”
Further, a label j1 is assigned to the input image pattern (normalized character pattern) X1, X2,..., XNj as a pattern used to create the binary mask pattern Mj1.
The above processing is performed for all character categories to be newly created.
[0069]
(Step 2a)
For all input image patterns Xt to be newly created, similarity SH (Xt, Mjk) is calculated with all binary mask patterns Mjk of all character categories, and Mlh having the maximum similarity is obtained. Now, it is assumed that the label jk is already assigned to the input image pattern Xt.
If l = j, it is determined that Xt has been correctly recognized in the jth character category, and a new label jh is assigned. Otherwise, Xt determines that it has been erroneously recognized and assigns a new label je.
After completion of this process, if there is no Xt that determines that all input image patterns Xt are erroneously recognized, the process proceeds to step 4a. Otherwise, go to step 3a.
[0070]
(Step 3a)
For the j-th character category to be newly created, h = 1.
For all input image patterns to which the label jk is assigned, if one or more input image patterns exist, a new binary mask pattern Mjh is created by the same procedure as in step 1a, and h = h + 1 is set.
This determination process is performed for all Kj labels of the jth character category.
Next, if there are one or more input image patterns to which the label je is assigned, the temporary binary mask pattern Mje is set for all the input image patterns Xt in the same procedure as in step 1a. The similarity SH (Xt, Mje) is calculated between all Xt and Mje, and the input image pattern XS having the maximum similarity is set as a new binary mask pattern Mjh, and h = h + 1 is set.
The number of new binary mask patterns for the jth character category is set as Kj = h.
This process is repeated for all character categories to be newly created, and a new binary mask pattern is stored in the binary mask pattern storage circuit 3. Then, the process returns to step 2a.
[0071]
(Step 4a)
A similarity SH (Xt, Mjk) is calculated between each binary mask pattern Mjk and all input image patterns Xt other than the jth character category, and a maximum similarity tjk for each binary mask pattern Mjk is obtained. This maximum similarity is assigned to the identification threshold value tjk of the binary mask pattern Mjk and stored in the binary mask pattern storage circuit 3. When these processes are completed for all binary mask patterns, the process ends.
[0072]
<Example 3>
This is an example in which the similarity SH is used in the same manner as in the second embodiment, but the binary mask pattern and the identification threshold are updated using the image pattern of the misrecognized character or the unlearned character.
First, information regarding the character category to be updated (referred to as the jth character category) is input from the image input / output device 6 to the binary mask pattern generation circuit 5. The binary mask pattern creation circuit 5 performs a binary mask on the input image for each character category of the character categories j = 1, 2,... Update the pattern.
[0073]
(Step 1b)
Let Kj be the number of binary mask patterns of the jth character category before update.
For all input image patterns Xt to be updated, the similarity SH (Xt, Xj) with all binary mask patterns Mjk (j = 1, 2,..., J) (k = 1, 2,..., Kj) of all character categories. Mjk) is calculated to obtain Mlh having the maximum similarity. Assume that the input image pattern Xt belongs to the jth character category.
If l ≠ j, it is determined that Xt is erroneously recognized, and a new label je is assigned.
For all character categories to be updated, after this process is completed, if there is no Xt that determines that all input image patterns Xt are erroneously recognized, the process proceeds to step 3b. Otherwise, go to step 2b.
[0074]
(Step 2b)
For all the input image patterns Xt to which the label je is assigned, a temporary binary mask pattern Mje is created in the same procedure as in step 1a of the second embodiment, and the similarity SH ( Xt, Mje) is calculated, and the input image pattern Xs having the maximum similarity is stored in the binary mask pattern storage circuit 3 as a binary mask pattern Mj (Kj + 1) to be added. Further, Kj = Kj + 1 is set, and the process returns to Step 1b.
[0075]
(Step 3b)
If each binary mask pattern Mjk and all input image patterns Xt other than the j-th character category are present, a similarity SH (Xt, Mjk) is calculated between them and the maximum similarity for each binary mask pattern Mjk. The degree t′jk is obtained, and when the maximum similarity is larger than the identification threshold value tjk (tjk = −n when newly created) stored in the binary mask pattern storage circuit 3, The identification threshold value tjk = t′jk of the value mask pattern Mjk is updated and the new value is stored in the binary mask pattern storage circuit 3. When these processes are completed for all binary mask patterns, the process ends.
[0076]
Here, in the second and third embodiments, the case where the similarity SH is used as the discrimination function has been described. However, the absolute value of the similarity SH can also be used as the discrimination function.
[0077]
In addition, as a method of creating a binary mask pattern, other methods based on known statistical inference (for example, “Statistics” written by Tetsuo Suzuki (Asakura Shoten) and “Pattern classification and scence analysis” R.O.Duda and the method by P.E.Hart (John Wiley & Sons Ins.) can also be used.
Also in the first embodiment, even when there is an image pattern that is not included in any character category, it is possible to create a binary mask pattern of a new character category using the image pattern. .
[0078]
Furthermore, in the above description, the characters are mainly shown, but general binary images such as graphics can be processed in the same manner.
As an example, a case where the present invention is applied to a method for recognizing a disease category based on a diagnosis result of multiple items used in medical treatment will be briefly described below.
For example, the results of a plurality of diagnostic items are represented on the check sheet as a binary image pattern as shown in FIG. 3 (for example, for a blood pressure item, “1” if the blood pressure is high, “0” if the blood pressure is low) The pattern on the check sheet is input as a binary image pattern and stored in advance based on typical diagnosis data of each disease name category. The category of the disease name is recognized by collating with the binary mask pattern that has been allowed to be used. As described above, after the recognition target is converted into a binary image and then recognized as in the above-described embodiment, the recognition performance of the disease name category can be improved.
[0079]
【The invention's effect】
As described above, according to the present invention, by comparing the input image pattern with the binary mask pattern stored in advance in the standard dictionary, the input image is identified, rejected, or recognized, so that an OCR user can make an error. In the operation of sequentially re-inspecting and correcting the recognition and reject pattern by visual inspection, it is not necessary to re-inspect the identified image, which has the advantage of reducing the work burden.
In the conventional character recognition method, one element of the recognition feature is often stored with a depth n (n> 1) bits, but in the present invention, each element of the binary mask pattern can be expressed with one bit. The storage capacity can be reduced to 1 / n. Furthermore, by using a known image coding method in combination, the storage capacity of the binary mask pattern can be further reduced, processing efficiency is high, and the dictionary capacity is compact.
[0080]
In addition, according to the present invention, it is not necessary to extract the feature amount related to the character line structure such as the direction of the character line and the connection relationship, so that the feature extraction process is simplified, and further, blurring, crushing, noise of the character background For example, there is an advantage that even if these character structures are deformed, the recognition performance is hardly affected. In addition to fading, blurring, and noise in the background of the text, even if a binary mask pattern created from an image pattern composed of the text as a black spot and the background as a white spot is used, the text is configured as a white spot and the background as a black spot. This is advantageous in that it has high recognition performance against blurring, crushing, noise in the background of characters, etc.
[0081]
Furthermore, according to the present invention, the input image pattern is compared with the binary mask pattern stored in advance in the standard dictionary to identify, reject, or recognize the input image, so that the OCR user can recognize and reject In the operation of sequentially re-inspecting and correcting the image, it is not necessary to re-inspect the recognized image, and there is an advantage that the work load is reduced.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating a configuration example of an image pattern identification / recognition processing apparatus according to the present invention.
FIG. 2 is a flowchart of an example of identification / rejection / recognition determination processing according to the present invention;
FIG. 3 is a diagram illustrating a pattern example for explaining specific determination processing of identification / rejection / recognition.
[Explanation of symbols]
1 binary image pattern storage circuit
2 Normalization processing circuit
3 binary mask pattern storage circuit
4 recognition judgment circuit
5 Binary mask pattern creation circuit
6 Image input / output device

Claims (5)

This is an image pattern identification / recognition method in which an input binary image pattern is collated with each binary mask pattern for each recognition category that has been created in advance, and a recognition category to which the binary image pattern belongs is determined. And
black pixel of n pixels plurality of the at a predetermined position the binary image pattern X black screen containing the number A and the binary mask pattern in the pattern constituted by Mjk (j is the number of categories, k is the number of mask patterns) When counting the number Bjk, a first number Cjk (hereinafter referred to as a first coincidence) is calculated in which the pixel value Xi of the binary image pattern X and the pixel value mi of the binary mask pattern are both black pixels. Steps,
A second step of calculating a number Djk (hereinafter referred to as a first degree of mismatch) in which the pixel value Xi of the binary image pattern X is a black pixel and the pixel value mi of the binary mask pattern is a white pixel;
By subtracting the first degree of coincidence calculated in the first step and the first degree of inconsistency calculated in the second step from the number of pixels n of the pattern, at the plurality of predetermined positions. A third step of calculating a number of white pixels in the binary image pattern and the binary mask pattern (hereinafter, second matching degree);
A number by which the binary image pattern becomes a white pixel and the binary mask pattern becomes a black pixel at the plurality of predetermined positions by subtracting the first coincidence calculated in the first step from B jk. (hereinafter, a second mismatch degree) and a fourth step of calculating,
A fifth step of calculating a third degree of coincidence by multiplying the first degree of coincidence and the second degree of coincidence;
A sixth step of calculating a third mismatch degree by multiplying the first mismatch degree and the second mismatch degree;
A first step configured by a seventh step of obtaining a similarity between the binary image pattern and the binary mask pattern by normalizing a value obtained by subtracting the third mismatch from the third match. Procedure and
A second procedure for recognizing that the binary image pattern belongs to a recognition category having the maximum similarity calculated by the first procedure;
An image pattern identification / recognition method characterized by comprising:   2. The image pattern identification / recognition method according to claim 1, wherein all positions in the pattern are used as a plurality of predetermined positions in the pattern. Has a procedure to set a threshold value in advance the every each binary mask pattern of each recognition category, instead of the second step,
When the similarity in the first procedure between the binary image pattern and the binary mask pattern of the recognition category is equal to or higher than the threshold value of the binary mask pattern for only one recognition category. ,
2. The image pattern identification / recognition method according to claim 1, wherein the binary image pattern has a third procedure for identifying the binary image pattern in its recognition category. The Zokuse not to the third one of the recognized category the input binary image pattern by the procedure, and two or more of the binary image patterns for recognition category and the binary mask pattern of the recognition categories If the degree of similarity in the first steps of equal to or greater than a threshold value of the binary mask pattern, according to claim 3, characterized in that a fourth procedure for rejecting said input binary image pattern Image pattern identification and recognition method. A fifth procedure for newly creating, adding, or updating a binary mask pattern by iterative learning using a predetermined similarity based on information on a black portion or a white portion of the input binary image pattern is provided. The image pattern identification / recognition method according to any one of claims 1 to 4.

Images