JP4469566B2 - Character recognition method - Google Patents

Description

上式（式１）、（式２）からエッジ方向θ（ｉ，ｊ）及びエッジ強度Ｖ（ｉ，ｊ）は次式のように表される物理量で定義される。
【００２３】
θ（ｉ，ｊ）＝tan^-1｛ｆｙ（ｉ，ｊ）／ｆｘ（ｉ，ｊ）｝ ・・・（式３）
Ｖ（ｉ，ｊ）＝｛ｆｘ（ｉ，ｊ）²＋ｆｙ（ｉ，ｊ）²｝^1/2 ・・・（式４）
ここで、ｆ（ｉ，ｊ）は座標（ｉ，ｊ）における画像の明るさであり、１が黒（文字部）であり、０が白（背景部）である。なお、２値化された文字パターンがメモリ上にビットマップ展開された場合には、上記ｉ，ｊはメモリのアドレスに相当することとなる。
【００２４】
図７、図８は図３に示す認識対象文字のエッジ特徴を抽出した結果の例であり、図７はエッジ方向を、図８はエッジ強度をそれぞれ示している。図９はエッジ特徴と方向コードの説明及びエッジ特徴テーブルの構成例を示す図である。ここで、エッジ特徴テーブルは前述したプログラム格納メモリ（又は、ハードディスク）に格納し、制御部５１によって適時読み出して実行制御するように構成している。 エッジ方向θ（ｉ，ｊ）は、図９（ａ）に示すように背景部から文字部に向かう方向を示しており、コンピュータ処理に供するべく、図９（ｂ）に示すような４つの方向コードに量子化している。また、方向が不定の場合は方向コードを５としている。また、エッジ強度Ｖ（ｉ，ｊ）は８段階に量子化しており、エッジ方向が不定の場合はエッジ強度を４としている。図９（ｃ）はエッジ特徴テーブルの構成例であり、特徴点の座標（ｉ，ｊ）、４つの方向コードに量子化されたエッジ方向θ、８段階に量子化されたエッジ強度Ｖで構成されている。
【００２５】
ステップＳ４：（エッジ間密度抽出）
次に、上記ステップＳ３で得たエッジ特徴から各走査方向に対して垂直方向に文字パターンのエッジ間密度を抽出する。ここで、エッジ間密度とは、各走査方向に対する垂直方向のエッジの相対関係を表す。エッジ間密度は、例えば、方向別に文字パターンを走査し、エッジ間の間隔から抽出することができる。
【００２６】
図１０は、エッジ間密度、分割位置情報抽出例及び分割位置情報テーブルの構成例を示す図である。ここで、分割位置情報テーブルは前述したプログラム格納メモリ（又は、ハードディスク）に格納し、制御部５１によって適時読み出して実行制御するように構成している。
【００２７】
実施例では、エッジ間密度はエッジ間の間隔から抽出し、上記ステップＳ２の外郭特徴抽出での各走査方向に対して垂直な方向について、エッジ間密度を求めエッジ間密度ヒストグラムを作成している。ここで、ヒストグラムには度数分布を表す表の意味と長方形の面積が量を表す意味とがあるが、本明細書では長方形の面積が量を表す意味で用いている。具体的には、走査方向ごとにエッジ間密度の量（物理量）を面積で表している。
【００２８】
図１０（ａ）は図３に示す認識対象文字のエッジ間密度を抽出した結果の一例であり、上記ステップＳ２の外郭特徴抽出での図６（ａ）に示す走査方向２、走査方向６に対して垂直な方向α（図１０（ｂ）参照）、及び走査方向０、走査方向４に対して垂直な方向β（図１０（ｂ）参照）、のエッジ間密度抽出の例である。ここで、符号２１は走査方向０、走査方向４に対して垂直方向のエッジ間密度ヒストグラムを示し、符号２２は走査方向２、走査方向６に対して垂直方向のエッジ間密度ヒストグラムを示す。これらのヒストグラムは２値化されてメモリに保存され、以下のステップで制御部５１によって適時読み出され、使用される。
【００２９】
ステップＳ５：（分割位置情報抽出）
次に、上記ステップＳ１で得た矩形情報、及び上記ステップＳ４で得たエッジ間密度に応じて、各走査方向およびそれぞれの走査方向に垂直な方向に前記２値化文字パターンに外接する矩形を垂直方向に分割数Ｎ個に分割する分割枠を決定し、矩形内の該垂直方向の各座標のそれぞれについて各座標ラインの分割位置及び座標ラインが分割位置を占めるエッジ間密度の割合を示す重みを抽出する。
【００３０】
例えば、分割数Ｎ＝８としたとき、図３に示す認識対象文字の水平方向の分割位置情報の抽出動作について説明すると、先ず、図１０（ａ）に示す上記ステップＳ４で得た水平方向のエッジ間密度ヒストグラム２２の総和を８等分するように分割して、図１０（ｂ）の符号２４に示すような分割枠を決定する。
次に、図４（ｂ）に示す上記ステップＳ１で得た縦横座標系での矩形１１内の水平方向座標それぞれについて、分割枠２４より座標ラインの分割位置を決定し、座標ラインのエッジ間密度ヒストグラムの内で決定された分割位置に割り当てられるエッジ間密度ヒストグラムと、水平方向の各分割位置のエッジ間密度ヒストグラムの総和より、各座標ラインの分割位置における重みを抽出する。
【００３１】
図１０（ｃ）は分割位置情報テーブルの構成例であり、水平方向の分割位置情報テーブルの例である。この例では水平方向の分割位置情報は、水平方向の座標ｘ、分割位置Ｐ、重みＷで構成されている。なお、この実施例では、重みを座標ラインが分割位置を占めるエッジ間密度の割合としたが、重みは任意に設定可能であり、各座標ラインの重みＷをＷ＝１として均等にしてもよい。
【００３２】
ステップＳ６：（方向別特徴合成）
次に、ステップＳ２での各走査方向に対して、下記（イ）〜（ロ）の手順により方向別特徴の合成を行う。
（イ）ステップＳ５で得た走査方向に対して垂直な方向の分割位置情報とステップＳ２で得た走査方向の外郭特徴を用いて、上記分割位置情報に含まれている各座標における外郭特徴の座標ラインの各外郭点でのエッジ方向とエッジ強度を上記ステップＳ３で得たエッジ特徴から求める。
（ロ）上記外郭点の次数情報及び方向と、分割位置情報に格納されている座標ラインの分割位置で決まる特徴格納位置を求める。
（ハ）上記（ロ）で求めたエッジ方向別のエッジ強度に対し上記分割位置情報に含まれている座標ラインの分割位置における重みを付加して累積する。
【００３３】
具体例として、図３に示した認識対象文字に関して図６（ａ）に示す走査方向６での特徴の合成について説明すると、図１０（ｃ）に示した走査方向６に対して垂直な方向（水平方向）の分割位置情報テーブル３２と図６（ｂ）に示した走査方向６の外郭特徴テーブル３０を用いて、分割位置情報テーブル３２に格納されている各座標における外郭特徴テーブル３０の座標ラインの各外郭点において、それぞれの外郭点におけるエッジ方向θとエッジ強度Ｖを上記ステップＳ３で得たエッジ特徴テーブル３１（図９（ｃ）参照）から得て、それら外郭点の次数Ｍ、走査方向６、分割位置情報テーブル３２に格納されている座標ラインの分割位置Ｐ、とで決まる特徴格納位置に対して、エッジ方向別にそのエッジ強度に分割位置情報テーブル３２に格納されている座標ラインの分割位置における重みＷを付加して累積する。
【００３４】
例えば、図１０（ｃ）の分割位置情報テーブル３２に格納されている水平方向の座標ｘ＝１における外郭特徴テーブル３０（図６（ｂ）参照）の座標ｘ＝１ラインの外郭点（１、３９）において、エッジ特徴テーブル３１の座標（１、３９）からエッジ方向θ＝１、エッジ強度Ｖ＝７を得て、外郭点（１、３９）の次数Ｍ＝１、走査方向Ｓ＝６、座標ｘ＝１ラインの分割位置Ｐ＝１、とで決まる特徴格納位置のエッジ方向θ＝１に、エッジ強度Ｖ＝７に座標ｘ＝１ラインの分割位置Ｐ＝１における前記重みＷ＝ｗ１１を付加して累積する。ここで、エッジ方向θが不定方向コードの場合（エッジ方向θ＝５の場合）は、走査方向Ｓと平行なエッジ方向に累積するものとする。
【００３５】
ステップＳ７：（特徴正規化）
次に、各特徴格納位置内でエッジ方向別に上記ステップＳ６で累積されたエッジ強度を正規化し、特徴量を得る。
あるエッジ方向ｉにおける正規化後の特徴量Ｄｉは正規化前の特徴量（エッジ強度の累積値）をＥＤｉとすると、次式で表される。
Ｄｉ＝（ＥＤｉ／ｍａｘ_ｊ（ＥＤｊ）） （式５）
ここで、ＥＤｊは特徴量（ベクトル）を表し、例えば、エッジ方向コード数が４個の場合ｊ＝１〜４となる。
また、式（５）の分母は４つのエッジ方向のうち最大の値（エッジ強度の累積値）を持つエッジ方向を選択することを意味する。（即ち、式（５）では、各特徴格納位置内において最大のエッジ方向が１になるように正規化することを意味する）。
【００３６】
ステップＳ８：（文字識別）
最後に、上記ステップＳ７で得た特徴量と予め用意した各文字の辞書テンプレートとのマッチングをとり文字パターンの分類を行う。
【００３７】
前記特許文献１記載の発明の文字認識方法では、分割枠は矩形を均等に分割して固定していたのに対し、本発明では、図１０（ｂ）に示すように、エッジ間密度により分割枠を変動させるので、図１７（ｂ）に示すような文字のはらいが長い場合などの文字の形状の変形を吸収した特徴の抽出を可能とする。
【００３８】
＜実施例２＞
本実施例では、図１７（ｇ）に示すような右上がり文字などの傾いた文字に対しても、特徴を安定的に求める。
図１１は文字認識方法の一実施例を示すフローチャートである。また、実施例においては、右上がり文字についての対応を説明する。また、図１１で、ステップＴ１〜Ｔ６は前述した図２のステップＳ１〜Ｓ６と同様な動作を行うので、以下、ステップＴ７以降の動作について説明する。
【００３９】
ステップＴ７（傾き検出）
ここでは、ステップＴ３（＝図２のステップＳ３のエッジ特徴抽出処理）で得たエッジ特徴（文字パターンそのものから直接得る特徴）、或いは、ステップＴ６（＝図２のステップＳ６の方向別特徴合成処理）で全方向に対して特徴を合成しエッジ方向別にエッジ強度を累積した累積エッジ強度（文字の変形を吸収した特徴）より、文字パターンの傾きを検出する。
【００４０】
例えば、ステップＴ３（＝図２のステップＳ３のエッジ特徴抽出処理）で得たエッジ特徴をエッジ方向別にカウントし、エッジ方向が右下がり方向（文字の輪郭線が右上がり方向）のカウント数の全体に占める割合（割合Ａ）、及び、エッジ方向が縦方向（文字の輪郭線が横方向）のカウント数の全体に占める割合（割合Ｂ）、及び、割合Ａと割合Ｂの比率（比率Ｃ）から、文字パターンが右上がり文字であるかどうか所定の右上がり判定条件で判定し、右上がり文字であると判定された場合において、上記比率Ｃより傾きを検出する。ここで、前述したように、エッジ特徴の方向は文字の輪郭線に対して垂直な方向を示している。
【００４１】
ステップＴ８（特徴正規化）
次に、ステップＴ６（＝図２のステップＳ６の方向別特徴合成処理）での累積エッジ強度の正規化を行う前、或いは、正規化後において、上記ステップＴ７で得た傾きに応じて、累積エッジ強度（正規化後は特徴量）に重みを付加する。
【００４２】
例えば、右上がり文字の対応に関しては、累積エッジ強度の正規化を行う前に、前記ステップＴ７で得た傾きに応じて、各特徴格納位置内でエッジ方向が右下がり方向（文字の輪郭線が右上がり方向）の累積エッジ強度でエッジ方向が縦方向（文字の輪郭線が横方向）の累積エッジ強度に重みを付加する。
【００４３】
図１２は右上がり文字に対応した重み付けの説明図である。
例えば、エッジ方向ｊを図９（ｂ）に示すエッジ方向コード（ｊ＝１〜４）とし、ある特徴格納位置ｉ内のあるエッジ方向ｊの正規化前の累積エッジ強度をＥＤｉｊとし、前記ステップＴ７で右上がり文字であると判定された場合において検出した文字パターンの傾きをθ（図１２（ａ）に示すようにθは水平方向に対して反時計回りの方向の角度とする）とし、次の式６を用いて、傾きθに応じて、エッジ方向が右下がり方向の累積エッジ強度ＥＤｉ４でエッジ方向が縦方向の累積エッジ強度ＥＤｉ３に重みを付加する。
ＥＤｉ３ ＝ ＥＤｉ３ ＋ ２ ＥＤｉ４ sinθ （式６）
なお、この方法は右上がり文字に関してのみ限定されるものではなく、全体に傾いた文字にも対応できるし、あらゆる傾きに対して対応可能である。また、その他に傾き文字の対応に関しては、座標系を文字の傾きに合わせて設定し、外郭特徴抽出の走査方向をその設定した座標系の座標軸方向として外郭特徴を抽出するようにしてもよい。この場合、右上がり文字の対応に関しては、例えば、図１２（ｂ）に示すように縦横座標系（ｘ軸、ｙ軸）を垂直方向と傾き方向として設定し、縦横方向の外郭特徴抽出は走査方向を縦横座標系の座標軸方向として外郭特徴を抽出するようにする。
【００４４】
＜変形例１＞
上記図２のフローチャートのステップＳ１（又は図１１のステップＴ１）の矩形情報抽出処理において、縦横座標系で文字パターンの矩形を抽出した後に、縦横座標系での矩形の２対角線を求め、各対角線に対して、斜め座標系を対角線方向とその対角線の垂直線方向の座標系として設定し、各ステップの処理対象方向を変更して、斜め方向の特徴抽出を縦横座標系での矩形の対角線の方向の特徴抽出とするように構成することにより、図１７（ｄ）、（ｅ）に示すように文字が縦長、横長の場合において、斜め方向の特徴を安定に求めることができる。
【００４５】
図１３は図３に示す認識対象文字の矩形情報抽出の一実施例であり、図１４はエッジ間密度抽出、分割位置情報抽出の一実施例である。
【００４６】
図２のステップＳ１（又は図１１のステップＴ１）で、まず、縦横座標系（ｘ軸、ｙ軸）での矩形１１を抽出した後に、矩形１１の２対角線（図１３（ａ）に示す対角線１３、対角線１４）を求め、対角線１３に対しての斜め座標系（ｒｘ１軸、ｒｙ１軸）のｒｘ１軸を対角線１３の方向、ｒｙ１軸を対角線１３の垂直線方向として設定し、その斜め座標系での矩形１５（図１３（ｂ））を抽出する。同様に、対角線１４に対しての斜め座標系（ｒｘ２軸、ｒｙ２軸）を設定し、矩形１６（図１３（ｃ））を抽出する。
【００４７】
次に、ステップＳ２（又は図１１のステップＴ２）で、斜め方向の外郭特徴抽出は走査方向を縦横座標系での矩形の対角線の方向として外郭特徴を抽出し、ステップＳ３（又は図１１のステップＴ３）では、前述したと同様に、文字パターンのエッジ特徴（エッジ方向及びエッジ強度）を抽出する。
【００４８】
また、ステップＳ４（又は図１１のステップＴ４）では、前述したと同様に、ステップＳ２（又は図１１のステップＴ２）の外郭特徴抽出での斜め走査方向に対して垂直な方向について、エッジ間密度（図１４の符号２５、２６で示す部分）を抽出する。
【００４９】
ステップＳ５（又は図１１のステップＴ５）では、ステップＳ１（又は図１１のステップＴ１）で得た斜め座標系での矩形情報、及びステップＳ４（又は図１１のステップＴ４）で得た斜め走査方向に対して垂直な方向のエッジ間密度に応じて、前述したと同様に、斜め走査方向に対して垂直な方向の分割位置（図１４の符号２７、２８で示す部分）情報を抽出する。
【００５０】
ステップＳ６（又は図１１のステップＴ６）は、斜め方向については、上記ステップＳ５（又は図１１のステップＴ５）で得た斜め走査方向に対して垂直な方向の分割位置情報、及び上記ステップＳ２（又は図１１のステップＴ２）で得た斜め走査方向の外郭特徴を用いて、前述したと同様に、分割位置情報に格納されている各座標における外郭特徴の座標ラインの各外郭点において、その外郭点におけるエッジ方向とエッジ強度を上記ステップＳ３（又は図１１のステップＴ３）で得たエッジ特徴から得て、外郭点の次数情報、方向、分割位置情報に格納されている座標ラインの分割位置、とで決まる特徴格納位置に対して、エッジ方向別にそのエッジ強度に分割位置情報に格納されている座標ラインの分割位置における重みを付加して累積する。
【００５１】
上記ステップＳ７では、前述したと同様に、各特徴格納位置内でエッジ方向別に上記ステップＳ６で累積されたエッジ強度を正規化し、特徴量を得る。
【００５２】
また、上記ステップＳ８では、上記ステップＳ７で得た特徴量と予め用意した各文字の辞書テンプレートとのマッチングをとり文字パターンの分類を行う。
【００５３】
このように、図２のステップＳ１で斜め方向の特徴抽出を縦横座標系で抽出した文字パターンの矩形の対角線の方向の特徴抽出とするように斜め座標系及び各ステップの処理対象方向を変更し、ステップＳ２〜Ｓ８を上述のように構成することにより、図１７（ｄ）、（ｅ）に示すように文字が縦長、横長の場合において、斜め方向の特徴を安定に求めることができる。
【００５４】
なお、上記実施例では、斜め座標系を対角線方向と対角線の垂直線方向の座標系としたが、斜め座標系を対角線方向ともう一方の対角線方向の座標系として設定し、各ステップの処理対象方向を変更して斜め方向の特徴抽出を行ってもよい。
【００５５】
＜変形例２＞
図１５は、図３に示す認識対象文字の斜め方向の分割位置情報の抽出についての説明図であり、図１５（ａ）は実施例１に関して、図１５（ｂ）は変形例１に関して本実施例を適応した場合の説明図である。
【００５６】
本実施例では、図２のステップＳ５（又は図１１のステップＴ５）の分割位置情報抽出動作において、斜め座標系での矩形（図１５（ａ）符号１２、図１５（ｂ）符号１５、１６）を分割する分割枠を決定するときに中央の分割線（符号４１、４２、４３、４４）を縦横座標系で抽出した矩形１１を２等分するように真ん中で区切った矩形２等分線（符号４５、４６、４７、４８）で固定し、他の分割枠については、上記矩形２等分線で２分割されたエッジ間密度ヒストグラムそれぞれを４等分（分割数Ｎ＝８の場合）するように分割して分割枠を決定する。このように構成しても斜め方向の特徴を安定に求めることができる。
【００５７】
＜変形例３＞
図１６は文字識別処理における斜め方向の特徴量の類似の度合いに対する重み付け方法の説明図である。
【００５８】
本実施例は、図２のステップＳ８（又は図１１のステップＴ９）の文字識別処理で、斜め方向の特徴量の類似の度合いのうち、縦横座標系での矩形の対角線付近（図１６（ａ）の斜線部分）からの距離に応じて特徴量の類似の度合いを弱く扱うようにする例であり、上記対角線付近の特徴は変動が大きいので重要度を下げて扱うようにする。本実施例の方法により、斜め方向の特徴を安定に求めることができる。
【００５９】
斜め方向の特徴量の類似の度合いの重み付けは、例えば、類似の度合いの計算前に、図１６（ｂ）に示すように、各対角線の中心を中心とした斜め４５°の線に関して左右対称な重み関数ｆ_ＷＲ（ｒｙ）、ｆ_ＷＬ（ｒｘ）を設定し（重み関数は任意に設定可能）、各分割枠の中央の座標と重み関数とで、分割枠の重みを決定しておき、類似の度合いの計算において、その分割枠に該当する特徴格納位置の特徴量の類似の度合いの計算時に分割枠の重みを付加した類似の度合いの計算を行うようにする。
【００６０】
例えば、類似の度合いの計算をユークリッド距離で行うと、次の式７により計算できる。
ｄ_ＤＥ＝｛Σ_ｉΣ_ｊ（Ｄ_ｉｊ−Ｅ_ｉｊ）^２｝^1/2 （式７）
ここで、Ｄ_ｉｊは文字パターンのある特徴格納位置ｉ内の特徴量（エッジ方向コード数４個の場合（ｊ＝１〜４））を表し、Ｅ_ｉｊは同様にある辞書テンプレートの特徴格納位置ｉ内の特徴量を表し、Σ_ｉは特徴格納位置ｉに関しての総和を表し、Σ_ｊはエッジ方向ｊに関しての総和を表す。
【００６１】
これに対して、分割数Ｎ＝８とし図１６（ｂ）に示すように各分割枠の重みＷをＷ_{ＲＭ 、}Ｗ_ＬＭ （Ｍ＝１、２、・・・、Ｎ）とすると、斜め方向の特徴量の重みを付加したユークリッド距離の計算は、右上がり方向（図６（ａ）に示す走査方向１、５）については次の式８、式１０により計算できる。
ｄ_ＤＥ１＝｛Σ_ｉΣ_ｊＷ_ＲＭ（Ｄ_ｉｊ−Ｅ_ｉｊ）^２｝^1/2 （式８）
また、左上がり方向（図６（ａ）に示す走査方向３、７）については次の式９、式１０により計算できる。
【００６２】
ｄ_ＤＥ２＝｛Σ_ｉΣ_ｊ Ｗ_ＬＭ（Ｄ_ｉｊ−Ｅ_ｉｊ）^２｝^1/2 （式９）
Ｍ ＝ ｍｏｄ（ｉ、Ｎ）＋１ （式１０）
ここで、Ｍは図１６（ｂ）で８分割で示される分割位置を示し、ｍｏｄ（ｉ、Ｎ）はｉをＮで割ったときの余りを表す。
【００６３】
また、対角線の中心を中心とした斜め４５°の線に関して左右対称な重み関数の設定の方法は、例えば、図１６（ｃ）に示すように、縦横座標系での矩形を斜め座標系で外接する矩形を求めて、重み関数の両端位置を決定することで行う。
【００６４】
本実施例の方法により、斜め方向の特徴を安定に求めることができる。
【００６５】
【発明の効果】
上記説明したように本発明の文字認識方法によれば、新たな辞書テンプレートの追加を行わずに、文字の形状の変形を吸収することができる。また、文字輪郭部分のノイズに影響されにくい。
【図面の簡単な説明】
【図１】本発明の文字認識方法を適用可能な文字認識装置の構成例を示すブロック図である。
【図２】文字認識方法の一実施例のフローチャートである。
【図３】認識対象文字の例を示す図である。
【図４】矩形情報の抽出例を示す図である。
【図５】外郭特徴の抽出例を示す図である。
【図６】走査方向、及び外郭特徴テーブルの構成例を示す図である。
【図７】エッジ特徴（エッジ方向）抽出例を示す図である。
【図８】エッジ特徴（エッジ強度）抽出例を示す図である。
【図９】エッジ特徴と方向コードの説明、及びエッジ特徴テーブルの構成例を示す図である。
【図１０】エッジ間密度抽出の例、分割位置情報抽出の例、及び分割位置情報テーブルの構成例を示す図である。
【図１１】文字認識方法の一実施例を示すフローチャートである。
【図１２】右上がり文字に対応した重み付けの説明図である。
【図１３】斜め方向の矩形情報の抽出例を示す図である。
【図１４】エッジ間密度抽出、分割位置情報の抽出例を示す図である。
【図１５】分割位置情報の抽出例を示す図である。
【図１６】文字識別処理における斜め方向の特徴量の類似の度合いに対する重み付け方法の説明図である。
【図１７】文字パターンの例を示す図である。
【符号の説明】
１１ 縦横座標系での矩形
１２ 斜め座標系での矩形
１３、１４ 縦横座標系での矩形の対角線
１５、１６ 斜め座標系での矩形
２１ 垂直方向のエッジ間密度ヒストグラム
２２ 水平方向のエッジ間密度ヒストグラム
２３ 垂直方向の分割枠
２４ 水平方向の分割枠
２５、２６ 斜め走査方向に対して垂直な方向のエッジ間密度ヒストグラム
２７、２８ 斜め走査方向に対して垂直な方向の分割枠
３０ 外郭特徴テーブル
３１ エッジ特徴テーブル
３２ 分割位置情報テーブル（分割位置情報）
４１、４２、４３、４４ 斜め座標系での矩形の中央の分割線
４５、４６、４７、４８ 縦横座標系での矩形の矩形２等分線
５０ 文字認識装置
５１ 制御部
５２ 読み取り部
５３ 文字切出し部
５４ 文字認識部
５５ 特徴抽出部
５６ 文字識別部
５７ 文字認識辞書
５８ 表示部[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a character recognition technique, and more particularly, to a character recognition technique for extracting a character feature from a binarized character pattern and recognizing the character using the feature.
[0002]
[Prior art]
In character recognition, an edge feature that determines the local direction of a character without the influence of noise and an outline feature that calculates a global feature of the character are synthesized, and the synthesized feature is matched with the dictionary template of each character to obtain a character pattern. There is a character recognition method having high recognition performance without being affected by noise in the character outline portion (see Patent Document 1).
[0003]
Here, the edge feature is a feature obtained by obtaining the direction and strength of the edge of the character pattern. The direction of the edge feature indicates a direction perpendicular to the outline of the character and does not vary greatly due to noise in the character outline. The outline feature is a feature obtained by obtaining a point (outline point) that changes from white (background) to black (character) when the character pattern is scanned from a plurality of directions, and represents a global feature of the character.
[0004]
[Patent Document 1]
Japanese Patent No. 3113769
[0005]
[Problems to be solved by the invention]
However, the conventional character recognition method has a problem in that the recognition rate may decrease when the shape of the character is deformed. For example, the recognition rate may decrease in the case where the character is long as shown in FIG. 17B or in the case where the character goes up to the right as shown in FIG.
[0006]
As a solution to this problem, for example, a method of increasing the number of character dictionary templates by predicting such deformation patterns of the character shape is conceivable, but it is impossible to predict all the deformation patterns. In addition, there are disadvantages such as an increase in the number of dictionary templates and a slower matching speed and an increased dictionary size.
[0007]
The present invention has been made to solve the above problems, and is a character recognition method that absorbs deformation of the shape of a character without adding a new dictionary template and is less susceptible to noise at the character outline portion. The purpose is to provide.
[0008]
  In order to solve the above-described problem, in the invention described in claim 1, a character reading unit that obtains binarized image data obtained by reading a document including a character to be recognized, and binarized image data are used. A character cutout unit that cuts out a binarized character pattern in units of one character, a feature extraction unit that extracts features from the binarized character pattern, and feature quantities extracted by the feature extraction unit are stored in a character recognition dictionary A character recognition method in a character recognition device having a character recognition unit that obtains a recognition result in comparison with a plurality of dictionary templates, wherein the feature extraction unit is a rectangle circumscribing a binarized character pattern in a vertical and horizontal coordinate system Rectangle information related to the character pattern, and45 °Rectangle information extraction step for extracting rectangle information related to each circumscribed rectangle in the rotated oblique coordinate system and storing it in the memory, sequentially extracting the outline points of the character pattern for each scanning direction, primary feature, secondary feature ,... Outline feature extraction step stored in memory as n-th feature, edge feature extraction step of extracting edge feature of character pattern and storing in memory, edge feature extracted in edge feature extraction stepEdge position calculated fromBased onEach in the outline feature extraction stepIn the direction perpendicular to the scanning directionAndWhile scanning the character pattern,Scan directionEach scan based on the inter-edge density extraction step in which the inter-edge density is extracted from the interval between the edges and stored in the memory, the rectangular information extracted in the rectangular information extraction step, and the inter-edge density extracted in the inter-edge density extraction step Determine the division frame that divides the rectangle circumscribing the character pattern in the direction perpendicular to the direction and the respective scanning direction in the vertical direction, and each coordinate in the vertical direction in the rectanglelineEach of the coordinates in which division framelineFind the dividing position ofFor each direction coordinate in the rectangle, the division position of each coordinate line and the coordinate line indicate the division position.OccupyEdge-to-edge densityA division position information acquisition step for obtaining a weight indicating a ratio and storing it in a memory, a feature synthesis step for performing feature synthesis for each scanning direction, and after the completion of feature synthesis, the accumulated edge strength is normalized to obtain the feature amount of the character pattern. A character recognition step for classifying the character pattern by matching the feature amount of the character pattern obtained in the feature amount acquisition step with the feature amount of the dictionary template of each character. The feature synthesis step includes, for each scanning direction, a vertical coordinate and outline feature extraction step included in the division position information in the direction perpendicular to the scanning direction obtained in the division position information acquisition step. Based on the obtained outline features for each scanning direction, an edge feature extraction step for the edge direction and edge strength at each outline point of the coordinate line of the outline feature at each coordinate A step of obtaining from the obtained edge feature, a step of obtaining a feature storage position on the memory determined by the degree information and direction of the outline point and the division position included in the division position information, and a step by edge direction obtained in the above step There is provided a character recognition method comprising the step of adding a weight in a divided frame corresponding to a vertical coordinate included in the divided position information and accumulating the edge strength.
  As a result, the edge feature that determines the local direction of the character without the influence of noise and the outline feature that determines the global feature of the character, the rectangle information of the character, the inter-edge density for each direction of the character extracted from the edge feature, Since the character pattern is classified on the basis of the synthesized features, the new dictionary template is added even when the character shape is deformed. Therefore, it is possible to absorb the deformation of the character shape and to have high recognition performance without being affected by the noise of the character outline portion.
[0009]
  According to a second aspect of the present invention, a character reading unit for obtaining binarized image data obtained by reading a document including a character to be recognized, and binarization for each character from the binarized image data. A character cutout unit that cuts out the converted character pattern, a feature extraction unit that extracts features from the binarized character pattern, and a plurality of dictionary templates in which the feature amounts extracted by the feature extraction unit are stored in the character recognition dictionary A character recognition method in a character recognition device having a character recognition unit that obtains a recognition result as compared with the above, wherein the feature extraction unit includes rectangular information related to a rectangle circumscribing a binarized character pattern in a vertical and horizontal coordinate system, and The vertical and horizontal coordinate systems are added to the character pattern.45 °Rectangle information extraction step for extracting rectangle information related to each circumscribed rectangle in the rotated oblique coordinate system and storing it in the memory, sequentially extracting the outline points of the character pattern for each scanning direction, primary feature, secondary feature ,... Outline feature extraction step stored in memory as n-th feature, edge feature extraction step of extracting edge feature of character pattern and storing in memory, edge feature extracted in edge feature extraction stepEdge position calculated fromBased onEach in the outline feature extraction stepIn the direction perpendicular to the scanning directionAndWhile scanning the character pattern,In the scanning directionEach scan based on the edge-to-edge density extraction step that extracts the edge-to-edge density from the interval between the edges and stores it in the memory, the rectangle information extracted at the rectangle information extraction step, and the edge-to-edge density extracted at the edge-to-edge density extraction step The division frame for dividing the rectangle circumscribing the character pattern in the direction and the direction perpendicular to each scanning direction is determined in the vertical direction, and in which division frame each coordinate in the vertical direction in the rectangle exists. Find the division position and each vertical coordinate in the rectanglelineEach of the coordinates in which division framelineFind the dividing position ofFor each direction coordinate in the rectangle, the division position of each coordinate line and the coordinate line indicate the division position.OccupyEdge-to-edge densityA division position information acquisition step for obtaining a weight indicating a ratio and storing it in the memory, a feature synthesis step for performing feature synthesis for each scanning direction, an inclination detection step for detecting the inclination of the character pattern after the feature synthesis is completed, and an inclination detection step A feature amount acquisition step of obtaining a feature amount of the character pattern normalized based on the accumulated edge strength according to the detected inclination, and the character recognition unit includes the feature amount of the character pattern obtained in the feature amount acquisition step. A character recognition step for classifying the character pattern by matching with the feature amount of the dictionary template for each character is provided. The feature synthesis step is performed for each scanning direction with respect to the scanning direction obtained in the division position information acquisition step. Each coordinate based on the vertical coordinates included in the vertical position division information and the outline features for each scanning direction obtained in the outline feature extraction step. Obtaining the edge direction and edge strength at each outline point of the outline line of the outline feature from the edge feature obtained in the edge feature extraction step, the degree information and direction of the outline point, and the division included in the division position information The step of obtaining the feature storage position in the memory determined by the position, and the weight in the divided frame corresponding to the vertical coordinate included in the divided position information is added to the edge strength for each edge direction obtained in the above step. And a step of accumulating the character.
  Thereby, in addition to the effect of the invention of the first aspect, it is possible to stably obtain characteristics even for a tilted character such as a right-upward character.
[0010]
  In the invention according to claim 3, the rectangular information extracting step includes extracting the rectangular information related to the rectangle circumscribing the character pattern in the vertical and horizontal coordinate system, and obtaining two diagonal lines of the rectangle; 3. The character recognition according to claim 1, further comprising a step of setting the diagonal coordinate system as a diagonal coordinate system and performing feature extraction in the diagonal direction of the rectangle as feature extraction in the diagonal direction. Provide a method. Thereby, the characteristic of the diagonal direction can be calculated | required stably.
[0011]
  In the invention according to claim 4, the division position information acquisition step determines the division pattern for dividing the rectangle of the character pattern in the oblique direction, and the character pattern obtained by extracting the central division line in the vertical and horizontal coordinate system. The character recognition method according to any one of claims 1 to 3, further comprising a step of fixing the rectangle by dividing it into two equal parts. Thereby, the characteristic of the diagonal direction can be calculated | required stably.
[0012]
  In the invention according to claim 5, the character recognizing unit obtains the divided position information of the feature storage position of the feature amount in the oblique direction and the character pattern extracted in the vertical and horizontal coordinate system with respect to the matching result of the feature amount in the oblique direction. The character recognition method according to claim 1, further comprising a step of weighting by applying a weight determined by a weight function that is symmetrical with respect to a diagonal 45 ° line centered on the center of a diagonal line of the rectangle. . Thereby, the characteristic of the diagonal direction can be calculated | required stably.
[0013]
DETAILED DESCRIPTION OF THE INVENTION
<Example 1>
FIG. 1 is a block diagram illustrating a configuration example of a character recognition device to which the character recognition method of the present invention can be applied. A character recognition device 50 includes a control unit 51, a reading unit 52, a character cutout unit 53, and a feature extraction unit 55. A character recognition unit 54, a character recognition dictionary 57, and a display unit 58.
[0014]
The control unit 51 includes a peripheral circuit (not shown) such as a CPU, a program storage memory, and a RAM, and controls each component of the character recognition device 50 described above and the overall operation of the character recognition device 50.
[0015]
The reading unit 52 reads a document including characters to be recognized, converts it into image data (binary data), and passes it to the character cutout unit 53. The character cutout unit 53 reads the original image obtained by the reading unit 52 in units of one character. A character image is cut out and passed to the feature extraction unit 55. The feature extraction unit 55 performs character feature extraction from the character image received from the character cutout unit 53 to obtain a feature amount. The character identification unit 56 compares the feature amount extracted by the feature extraction unit 55 with a plurality of dictionary templates stored in the character recognition dictionary 57 to obtain a recognition result. The display unit 58 displays a character image and a recognition result for one original.
[0016]
In the embodiment, the character recognition unit 54 including the character cutout unit 53, the feature extraction unit 55, and the character identification unit 56 is configured by a program, and the program storage memory (or hard disk) described above together with the character recognition control program. And the like are read out in a timely manner and controlled for execution.
[0017]
FIG. 2 is a flowchart of an embodiment of the character recognition method based on the present invention, which corresponds to the character recognition operation of the character recognition unit 54. FIG. 3 shows an example of a character to be recognized. In FIG. 3B, '·' represents a background portion (white) and '1' represents a character portion (black).
[0018]
Step S1: (rectangular information extraction)
  First, for the binarized character pattern, VerticalIn the abscissaRectangle information of the bounding rectangle,And rectangle information of a rectangle circumscribing in the oblique coordinate system is extracted. Here, the binarized character pattern means image data (binarized data) obtained by cutting out a character image in units of one character from the original image obtained by the reading unit 52 by the character cutting unit 53.
  FIG. 4 is an example in which rectangular information of the recognition target character shown in FIG. 3 is extracted. As shown in FIG. 4A, the x-axis of the vertical and horizontal coordinate system (x-axis, y-axis) is horizontal and the y-axis is vertical. DirectionThen, as shown in FIG. 4B, the rectangle information of the rectangle 11 in the vertical and horizontal coordinate system isWhen the oblique coordinate system (rx axis, ry axis) is a coordinate system obtained by rotating the vertical / horizontal coordinate system (x axis, y axis) by 45 degrees, as shown in FIG.SlantRectangle 12 in the coordinate systemRectangle informationIs extracted.
[0019]
Step S2: (Outline feature extraction)
Next, outline features of the character pattern binarized for each direction are extracted. The outline feature extraction can be performed by the same operation as the outline feature extraction process used in the well-known character recognition method using the outline feature. That is, a character pattern is scanned from a plurality of directions, and a point (outline point) that first crosses the character line in a certain scanning direction is a primary feature, and similarly, the second time, the third time,. The crossing points (outline points) are stored in the outline feature table as secondary features, tertiary features,..., N-th feature, respectively.
FIG. 5 is an example of the result of extracting the outline feature of the recognition target character shown in FIG. 3, and is an example of the outline feature extraction in the scanning direction 6 shown in FIG. FIG. 6 is a diagram illustrating a configuration example of the scanning direction and the outline feature table. FIG. 6A illustrates a configuration example of the scanning direction and FIG. 6B illustrates a configuration example of the outline feature table. Here, the outline feature table is stored in the above-described program storage memory (or hard disk), and is configured to be read out and controlled by the control unit 51 in a timely manner.
[0020]
In the example of FIG. 6, the outline feature includes a scanning direction S, coordinates (x, y) of an outline point, and an order M.
[0021]
Step S3: (Edge feature extraction)
Next, the edge characteristics (edge direction and edge strength) of the binarized character pattern are extracted. For example, the edge direction and the edge strength are extracted by the following method.
If the x-axis is the horizontal direction, the y-axis is the vertical direction, the coordinate in the x direction is i, and the coordinate in the y direction is j, the differential value fx (i, j) in the x direction and the differential value fy ( i, j) is expressed by the following expressions (Expression 1) and (Expression 2).
[0022]

From the above equations (Equation 1) and (Equation 2), the edge direction θ (i, j) and the edge strength V (i, j) are defined by physical quantities represented by the following equations.
[0023]
θ (i, j) = tan^-1{Fy (i, j) / fx (i, j)} (Expression 3)
V (i, j) = {fx (i, j)²+ Fy (i, j)²}^1/2   ... (Formula 4)
Here, f (i, j) is the brightness of the image at the coordinates (i, j), 1 is black (character part), and 0 is white (background part). When a binarized character pattern is developed on a memory as a bitmap, i and j correspond to memory addresses.
[0024]
7 and 8 are examples of the result of extracting the edge features of the recognition target character shown in FIG. 3, FIG. 7 shows the edge direction, and FIG. 8 shows the edge strength. FIG. 9 is a diagram illustrating an edge feature and a direction code and a configuration example of an edge feature table. Here, the edge feature table is stored in the above-described program storage memory (or hard disk), and is configured to be read out and controlled by the control unit 51 in a timely manner. The edge direction θ (i, j) indicates the direction from the background portion to the character portion as shown in FIG. 9A, and four directions as shown in FIG. 9B are used for computer processing. The code is quantized. When the direction is indefinite, the direction code is 5. Further, the edge strength V (i, j) is quantized in 8 steps, and the edge strength is set to 4 when the edge direction is indefinite. FIG. 9C shows an example of the configuration of the edge feature table, which is composed of feature point coordinates (i, j), edge direction θ quantized into four direction codes, and edge strength V quantized into eight levels. Has been.
[0025]
Step S4: (Density extraction between edges)
  Next, the inter-edge density of the character pattern is extracted in the direction perpendicular to each scanning direction from the edge feature obtained in step S3. Here, the edge-to-edge density represents a relative relationship between edges in the vertical direction with respect to each scanning direction. The edge density is, for example,By directionScan character pattern, spacing between edgesFromCan be extracted.
[0026]
FIG. 10 is a diagram illustrating a configuration example of an edge density, an example of division position information extraction, and a division position information table. Here, the division position information table is stored in the above-described program storage memory (or hard disk), and is configured to be read out and controlled by the control unit 51 in a timely manner.
[0027]
In the embodiment, the inter-edge density is extracted from the interval between the edges, and the inter-edge density is obtained in the direction perpendicular to each scanning direction in the outline feature extraction in step S2 to create an inter-edge density histogram. . Here, the histogram has a meaning of a table representing a frequency distribution and a meaning of a rectangular area representing a quantity, but in this specification, a rectangular area represents a quantity. Specifically, the amount (physical quantity) of the density between edges for each scanning direction is represented by an area.
[0028]
  FIG. 10A is an example of the result of extracting the inter-edge density of the recognition target character shown in FIG. 3, and in the scanning direction 2 and the scanning direction 6 shown in FIG. 6A in the outline feature extraction in step S2. Direction perpendicular toα (see FIG. 10B),And a direction perpendicular to the scanning direction 0 and the scanning direction 4β (see FIG. 10B),It is an example of density extraction between edges. Here, reference numeral 21 denotesFor scan direction 0, scan direction 4A vertical edge-to-edge density histogram is shown.Perpendicular to scanning direction 2 and scanning direction 6The edge-to-edge density histogram is shown. These histograms are binarized and stored in the memory, and are read and used by the control unit 51 in a timely manner in the following steps.
[0029]
Step S5: (Division position information extraction)
  Next, according to the rectangular information obtained in step S1 and the edge density obtained in step S4,Each scan direction and the direction perpendicular to each scan directionIn the binary character patternCircumscribeRectangleVerticallyDecide the division frame to be divided into N divisions,Each coordinate in the vertical directionFor each of the above, a division position of each coordinate line and a weight indicating a ratio of the density between edges that the coordinate line occupies the division position are extracted.
[0030]
  For example, when the division number N = 8, the horizontal division position information extraction operation for the recognition target character shown in FIG. 3 will be described. First, the horizontal direction obtained in step S4 shown in FIG. The total sum of the inter-edge density histogram 22 is divided into eight equal parts, and a division frame as shown by reference numeral 24 in FIG. 10B is determined.
  Next, for each horizontal coordinate in the rectangle 11 in the vertical and horizontal coordinate system obtained in step S1 shown in FIG. 4B, the division position of the coordinate line is determined from the division frame 24, and the inter-edge density of the coordinate line is determined. An edge-to-edge density histogram assigned to the determined split positions in the histogram;For each horizontal split positionEdge density histogramofThe weight at the division position of each coordinate line is extracted from the sum.
[0031]
FIG. 10C shows a configuration example of the division position information table, which is an example of the division position information table in the horizontal direction. In this example, the division position information in the horizontal direction is composed of a coordinate x in the horizontal direction, a division position P, and a weight W. In this embodiment, the weight is a ratio of the density between edges in which the coordinate line occupies the division position. However, the weight can be arbitrarily set, and the weight W of each coordinate line may be set equal to W = 1. .
[0032]
  Step S6: (Feature synthesis by direction)
  Next, for each scanning direction in step S2,The direction-specific features are synthesized according to the following procedures (a) to (b).
(I)Obtained in step S5scanningUsing the division position information in the direction perpendicular to the direction and the outline feature in the scanning direction obtained in step S2, the division position information isIncludedEach outline of the coordinate line of the outline feature at each coordinateIn pointFrom the edge feature obtained in step S3 above, the edge direction and edge strengthAskingThe
(B) AboveOrder information of outline pointsas well asdirectionAnd minutesThe division position of the coordinate line stored in the split position informationIn placeDetermined feature storage locationAsk.
(C) Obtained in (b) aboveBy edge directionofFor edge strengthFor the aboveIn division position informationIncludedA weight is added at the position where the coordinate line is divided to accumulate.
[0033]
  As a specific example, the composition of features in the scanning direction 6 shown in FIG. 6A with respect to the recognition target character shown in FIG. 3 will be described. The direction perpendicular to the scanning direction 6 shown in FIG. The coordinate line of the outline feature table 30 at each coordinate stored in the division position information table 32 using the division position information table 32 in the horizontal direction and the outline feature table 30 in the scanning direction 6 shown in FIG. Edge feature table 31 obtained in step S3 with the edge direction θ and edge strength V at each contour point.(See Fig. 9 (c))The feature storage position determined by the degree M of the outline points, the scanning direction 6, and the division position P of the coordinate line stored in the division position information table 32 is divided into edge strengths for each edge direction. A weight W at the division position of the coordinate line stored in the position information table 32 is added and accumulated.
[0034]
  For example,In FIG.Outline feature table 30 at horizontal coordinate x = 1 stored in division position information table 32(See FIG. 6 (b))At the contour point (1, 39) of the line x = 1, the edge direction θ = 1 and the edge strength V = 7 are obtained from the coordinates (1, 39) of the edge feature table 31, and the contour point (1, 39) is obtained. Of the feature storage position determined by the order M = 1, the scanning direction S = 6, the coordinate x = 1 line division position P = 1, the edge intensity V = 7 and the coordinate x = 1 line. The weight W = w11 at the division position P = 1 is added and accumulated. Here, when the edge direction θ is an indefinite direction code (when the edge direction θ = 5), accumulation is performed in the edge direction parallel to the scanning direction S.
[0035]
Step S7: (feature normalization)
Next, the edge strength accumulated in step S6 is normalized for each edge direction within each feature storage position to obtain a feature amount.
The normalized feature quantity Di in a certain edge direction i is expressed by the following expression, where EDi is the feature quantity before normalization (cumulative value of edge strength).
Di = (EDi / max_j(EDj)) (Formula 5)
Here, EDj represents a feature quantity (vector). For example, when the number of edge direction codes is 4, j = 1 to 4.
Further, the denominator of the equation (5) means that the edge direction having the maximum value (the accumulated value of the edge strength) is selected from the four edge directions. (That is, the expression (5) means normalization so that the maximum edge direction is 1 in each feature storage position).
[0036]
Step S8: (Character identification)
Finally, the character pattern is classified by matching the feature amount obtained in step S7 with a dictionary template for each character prepared in advance.
[0037]
In the character recognition method of the invention described in Patent Document 1, the dividing frame is divided and fixed in a uniform rectangle, whereas in the present invention, as shown in FIG. Since the frame is fluctuated, it is possible to extract a feature that absorbs the deformation of the character shape, such as when the character has a long stroke as shown in FIG.
[0038]
<Example 2>
In the present embodiment, the feature is stably obtained even for a tilted character such as a right-upward character as shown in FIG.
FIG. 11 is a flowchart showing an embodiment of the character recognition method. Further, in the embodiment, a description will be given of correspondence with respect to a right-upward character. In FIG. 11, steps T1 to T6 perform the same operations as steps S1 to S6 of FIG. 2 described above, so the operations after step T7 will be described below.
[0039]
Step T7 (inclination detection)
Here, the edge feature (feature obtained directly from the character pattern itself) obtained in step T3 (= edge feature extraction processing in step S3 in FIG. 2) or step T6 (= direction-specific feature synthesis processing in step S6 in FIG. 2). ), The inclination of the character pattern is detected from the accumulated edge strength (feature that absorbed the deformation of the character) obtained by combining the features in all directions and accumulating the edge strength in each edge direction.
[0040]
For example, the edge features obtained in step T3 (= the edge feature extraction process in step S3 in FIG. 2) are counted for each edge direction, and the total number of counts in which the edge direction is in the right-down direction (the character outline is in the right-up direction). Occupying ratio (ratio A), and the ratio (ratio B) of the total number of counts in which the edge direction is vertical (the outline of the character is horizontal), and the ratio of ratio A and ratio B (ratio C) Therefore, whether or not the character pattern is a right-upward character is determined by a predetermined right-up character determination condition. If it is determined that the character pattern is a right-upward character, the slope is detected from the ratio C. Here, as described above, the direction of the edge feature indicates a direction perpendicular to the outline of the character.
[0041]
Step T8 (feature normalization)
Next, before or after normalization of the accumulated edge intensity in step T6 (= direction-specific feature synthesis processing in step S6 in FIG. 2), the accumulated edge strength is accumulated according to the slope obtained in step T7. A weight is added to the edge strength (feature value after normalization).
[0042]
For example, with respect to the correspondence of a right-upward character, before normalizing the accumulated edge strength, the edge direction in the respective feature storage positions is changed to the right-down direction (the outline of the character is changed) according to the slope obtained in step T7. A weight is added to the accumulated edge strength in the vertical direction (the contour line of the character is in the horizontal direction) with the cumulative edge strength in the upward direction).
[0043]
FIG. 12 is an explanatory diagram of the weighting corresponding to the characters rising to the right.
For example, the edge direction j is the edge direction code (j = 1 to 4) shown in FIG. 9B, the cumulative edge strength before normalization of a certain edge direction j in a certain feature storage position i is EDij, and the step The inclination of the character pattern detected when it is determined that the character is a right-upward character at T7 is θ (as shown in FIG. 12A, θ is an angle in the counterclockwise direction with respect to the horizontal direction). Using the following expression 6, weights are added to the accumulated edge strength EDi4 whose edge direction is the right-down direction and the accumulated edge strength EDi3 whose edge direction is the vertical direction according to the inclination θ.
EDi3 = EDi3 + 2 EDi4 sinθ (Formula 6)
Note that this method is not limited to characters that rise to the right, and can handle characters that are slanted as a whole, and can handle any slant. In addition, regarding the correspondence of the tilted character, the coordinate system may be set according to the tilt of the character, and the contour feature may be extracted using the scanning direction of the contour feature extraction as the coordinate axis direction of the set coordinate system. In this case, with respect to the correspondence of the right-upward character, for example, as shown in FIG. 12B, the vertical and horizontal coordinate systems (x-axis and y-axis) are set as the vertical direction and the inclination direction, and the outline feature extraction in the vertical and horizontal directions is scanned. The outline feature is extracted with the direction as the coordinate axis direction of the vertical / horizontal coordinate system.
[0044]
<Modification 1>
In the rectangular information extraction process in step S1 (or step T1 in FIG. 11) of the flowchart of FIG. 2, after extracting the rectangle of the character pattern in the vertical and horizontal coordinate system, two diagonal lines in the vertical and horizontal coordinate system are obtained and each diagonal line is obtained. On the other hand, the diagonal coordinate system is set as the coordinate system of the diagonal direction and the vertical line direction of the diagonal line, the processing target direction of each step is changed, and the feature extraction in the diagonal direction is performed for the rectangular diagonal line in the vertical and horizontal coordinate system. By configuring so that direction feature extraction is performed, it is possible to stably obtain oblique direction features when the characters are vertically long and horizontally long as shown in FIGS.
[0045]
FIG. 13 shows an example of rectangular information extraction of the recognition target character shown in FIG. 3, and FIG. 14 shows an example of edge-to-edge density extraction and division position information extraction.
[0046]
In step S1 of FIG. 2 (or step T1 of FIG. 11), first, the rectangle 11 in the vertical and horizontal coordinate system (x axis, y axis) is extracted, and then two diagonal lines of the rectangle 11 (diagonal lines shown in FIG. 13A). 13, the diagonal line 14), the rx1 axis of the diagonal coordinate system (rx1 axis, ry1 axis) with respect to the diagonal line 13 is set as the direction of the diagonal line 13, and the ry1 axis is set as the vertical line direction of the diagonal line 13. The rectangle 15 at (FIG. 13B) is extracted. Similarly, an oblique coordinate system (rx2 axis, ry2 axis) with respect to the diagonal line 14 is set, and a rectangle 16 (FIG. 13C) is extracted.
[0047]
Next, in step S2 (or step T2 in FIG. 11), in the outline feature extraction in the oblique direction, the outline feature is extracted with the scanning direction as the direction of the diagonal line of the rectangle in the vertical and horizontal coordinate system, and step S3 (or step in FIG. 11). In T3), as described above, the edge features (edge direction and edge strength) of the character pattern are extracted.
[0048]
In step S4 (or step T4 in FIG. 11), as described above, the edge-to-edge density in the direction perpendicular to the oblique scanning direction in the outline feature extraction in step S2 (or step T2 in FIG. 11). (Portions indicated by reference numerals 25 and 26 in FIG. 14) are extracted.
[0049]
In step S5 (or step T5 in FIG. 11), the rectangular information in the oblique coordinate system obtained in step S1 (or step T1 in FIG. 11) and the oblique scanning direction obtained in step S4 (or step T4 in FIG. 11). In the same manner as described above, division position information (parts indicated by reference numerals 27 and 28 in FIG. 14) in the direction perpendicular to the oblique scanning direction is extracted in accordance with the edge-to-edge density.
[0050]
In step S6 (or step T6 in FIG. 11), in the oblique direction, the division position information in the direction perpendicular to the oblique scanning direction obtained in step S5 (or step T5 in FIG. 11) and the above step S2 ( Alternatively, using the outline feature in the oblique scanning direction obtained in step T2) in FIG. 11, as described above, at each outline point of the outline feature coordinate line at each coordinate stored in the division position information, the outline is obtained. The edge direction and the edge strength at the point are obtained from the edge feature obtained in step S3 (or step T3 in FIG. 11), and the coordinate line division position stored in the degree information, direction, and division position information of the outline point, The feature storage position determined by and is added to the edge strength for each edge direction by adding the weight at the division position of the coordinate line stored in the division position information. To.
[0051]
In step S7, as described above, the edge strength accumulated in step S6 is normalized for each edge direction in each feature storage position to obtain a feature amount.
[0052]
In step S8, the character pattern is classified by matching the feature amount obtained in step S7 with a dictionary template for each character prepared in advance.
[0053]
As described above, the oblique coordinate system and the processing target direction of each step are changed so that the feature extraction in the diagonal direction in step S1 in FIG. 2 is the feature extraction in the diagonal direction of the rectangle of the character pattern extracted in the vertical and horizontal coordinate system. By configuring steps S2 to S8 as described above, it is possible to stably obtain the characteristic in the oblique direction when the characters are vertically long and horizontally long as shown in FIGS. 17 (d) and 17 (e).
[0054]
In the above embodiment, the diagonal coordinate system is a diagonal coordinate system and a diagonal vertical coordinate system, but the diagonal coordinate system is set as a diagonal coordinate system and the other diagonal coordinate system, and the processing target of each step The feature extraction in the oblique direction may be performed by changing the direction.
[0055]
<Modification 2>
FIG. 15 is an explanatory diagram regarding the extraction of the division position information in the oblique direction of the recognition target character shown in FIG. 3, FIG. 15 (a) relates to the first embodiment, and FIG. 15 (b) relates to the first modification. It is explanatory drawing at the time of applying an example.
[0056]
In this embodiment, in the division position information extraction operation in step S5 in FIG. 2 (or step T5 in FIG. 11), the rectangles in the oblique coordinate system (reference numeral 12 in FIG. 15 and reference numerals 15 and 16 in FIG. 15B). A rectangular bisector obtained by dividing the middle dividing line (reference numerals 41, 42, 43, 44) in the vertical / horizontal coordinate system into two equal parts when determining the dividing frame for dividing the middle part). (The reference numerals 45, 46, 47, and 48). For the other divided frames, each of the edge density histograms divided into two by the rectangular bisector is divided into four equal parts (when the number of divisions N = 8). The division frame is determined by dividing as described above. Even in such a configuration, it is possible to stably obtain the characteristic in the oblique direction.
[0057]
<Modification 3>
FIG. 16 is an explanatory diagram of a weighting method for the degree of similarity of feature amounts in the oblique direction in the character identification process.
[0058]
In the present embodiment, in the character identification processing in step S8 in FIG. 2 (or step T9 in FIG. 11), among the similarities of the feature quantities in the oblique direction, the vicinity of the diagonal of the rectangle in the vertical and horizontal coordinate system (FIG. This is an example in which the degree of similarity of the feature quantity is handled weakly according to the distance from the hatched portion)), and the feature in the vicinity of the diagonal line has a large fluctuation, so that the importance is lowered. By the method of the present embodiment, the characteristic in the oblique direction can be obtained stably.
[0059]
For example, as shown in FIG. 16B, the weighting of the similarity degree of the feature amount in the oblique direction is bilaterally symmetric with respect to a diagonal line of 45 ° with the center of each diagonal line as shown in FIG. Weight function f_WR(Ry), f_WL(Rx) is set (the weighting function can be arbitrarily set), the weight of the divided frame is determined based on the coordinates of the center of each divided frame and the weighting function. At the time of calculating the degree of similarity of the feature quantity at the corresponding feature storage position, the degree of similarity is calculated by adding the weight of the dividing frame.
[0060]
For example, if the degree of similarity is calculated with the Euclidean distance, it can be calculated by the following equation (7).
d_DE= {Σ_iΣ_j(D_ij-E_ij)²}^1/2              (Formula 7)
Where D_ijRepresents a feature amount in a feature storage position i having a character pattern (when the number of edge direction codes is four (j = 1 to 4)), E_ijRepresents the feature amount in the feature storage position i of a dictionary template, and Σ_iRepresents the sum for the feature storage position i, and Σ_jRepresents the sum in the edge direction j.
[0061]
On the other hand, the division number N = 8 and the weight W of each division frame is set to W as shown in FIG._RM  ,W_LM If (M = 1, 2,..., N), the calculation of the Euclidean distance with the weight of the feature amount in the oblique direction is performed in the upward direction (scanning directions 1 and 5 shown in FIG. 6A). Can be calculated by the following equations 8 and 10.
d_DE1= {Σ_iΣ_jW_RM(D_ij-E_ij)²}^1/2        (Formula 8)
Further, the upward left direction (scanning directions 3 and 7 shown in FIG. 6A) can be calculated by the following formulas 9 and 10.
[0062]
d_DE2= {Σ_iΣ_j W_LM(D_ij-E_ij)²}^1/2        (Formula 9)
M = mod (i, N) +1 (Equation 10)
Here, M represents a division position indicated by 8 divisions in FIG. 16B, and mod (i, N) represents a remainder when i is divided by N.
[0063]
In addition, for example, as shown in FIG. 16C, a method of setting a weight function that is symmetric with respect to a diagonal line of 45 ° centered on the center of the diagonal line is used to circumscribe a rectangle in the vertical and horizontal coordinate system in the diagonal coordinate system. This is done by obtaining a rectangle to be determined and determining both end positions of the weight function.
[0064]
By the method of the present embodiment, the characteristic in the oblique direction can be obtained stably.
[0065]
【The invention's effect】
As described above, according to the character recognition method of the present invention, it is possible to absorb the deformation of the character shape without adding a new dictionary template. Moreover, it is hard to be influenced by the noise of a character outline part.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration example of a character recognition apparatus to which a character recognition method of the present invention can be applied.
FIG. 2 is a flowchart of an embodiment of a character recognition method.
FIG. 3 is a diagram illustrating an example of recognition target characters.
FIG. 4 is a diagram illustrating an example of extracting rectangular information.
FIG. 5 is a diagram illustrating an example of outline feature extraction;
FIG. 6 is a diagram illustrating a configuration example of a scanning direction and an outline feature table.
FIG. 7 is a diagram illustrating an example of edge feature (edge direction) extraction.
FIG. 8 is a diagram showing an example of edge feature (edge strength) extraction.
FIG. 9 is a diagram illustrating an edge feature and a direction code, and a configuration example of an edge feature table.
10 is a diagram illustrating an example of edge density extraction, an example of division position information extraction, and a configuration example of a division position information table. FIG.
FIG. 11 is a flowchart illustrating an embodiment of a character recognition method.
FIG. 12 is an explanatory diagram of weighting corresponding to a right-upward character.
FIG. 13 is a diagram illustrating an example of extracting rectangular information in an oblique direction.
FIG. 14 is a diagram illustrating an example of edge density extraction and division position information extraction.
FIG. 15 is a diagram illustrating an example of extraction of division position information.
FIG. 16 is an explanatory diagram of a weighting method for the degree of similarity of feature amounts in an oblique direction in character identification processing.
FIG. 17 is a diagram illustrating an example of a character pattern.
[Explanation of symbols]
11 Rectangle in ordinate and abscissa
12 Rectangle in oblique coordinate system
13, 14 Diagonal diagonal in ordinate and abscissa
15, 16 Rectangle in oblique coordinate system
21 Vertical edge-to-edge density histogram
22 Horizontal edge-to-edge density histogram
23 Vertical dividing frame
24 Horizontal dividing frame
25, 26 Inter-edge density histogram in a direction perpendicular to the oblique scanning direction
27, 28 Dividing frame in a direction perpendicular to the oblique scanning direction
30 outline feature table
31 Edge feature table
32 Dividing position information table (dividing position information)
41, 42, 43, 44 The dividing line at the center of the rectangle in the oblique coordinate system
45, 46, 47, 48 Rectangle rectangle bisector in ordinate and abscissa
50 character recognition device
51 Control unit
52 Reading unit
53 character cutout
54 Character recognition part
55 Feature extraction unit
56 Character identification part
57 Character recognition dictionary
58 display

Claims (5)

A character reading unit that obtains binarized image data obtained by reading a document including a character to be recognized, and a character cutout unit that cuts out a binarized character pattern from the binarized image data in units of one character. A feature extraction unit that extracts features from a binarized character pattern, and a character recognition unit that obtains a recognition result by comparing the feature amount extracted by the feature extraction unit with a plurality of dictionary templates stored in the character recognition dictionary A character recognition method in a character recognition device comprising:
The feature extraction unit includes:
Rectangle information relating to a rectangle circumscribing the binarized character pattern in the ordinate / horizontal coordinate system and rectangle information relating to the rectangle circumscribing in the oblique coordinate system obtained by rotating the ordinate / horizontal coordinate system by 45 ° to the character pattern are extracted. Rectangle information extraction step to be stored in the memory,
An outline feature extraction step for sequentially extracting outline points of the character pattern for each scanning direction and storing them in a memory as primary features, secondary features,..., N-th feature;
An edge feature extraction step of extracting an edge feature of the character pattern and storing it in a memory;
For the direction perpendicular to the scanning direction in the outer feature extraction step based on the edge position obtained from the extracted edge feature at the edge feature extraction step in the vertical direction while scanning a pre Symbol character pattern An inter-edge density extraction step of extracting an inter-edge density from an interval between edges in the scanning direction and storing it in a memory;
Based on the rectangular information extracted in the rectangular information extracting step and the inter-edge density extracted in the inter-edge density extracting step, a rectangle circumscribing the character pattern in each scanning direction and in a direction perpendicular to each scanning direction is vertically determining the divided frame is divided into, determine the division position of the or respective coordinate lines each coordinate line of the vertical direction of the rectangle is present which partition frame, dividing each coordinate line for each direction coordinates in the rectangular A division position information acquisition step of obtaining a weight indicating a ratio of the density between edges in which the position and the coordinate line occupy the division position and storing the weight in the memory
A feature synthesis step for performing feature synthesis for each of the scanning directions;
After the feature synthesis is completed, a feature amount acquisition step for obtaining the feature amount of the character pattern by normalizing the accumulated edge strength;
With
The character recognizing unit includes a character recognition step of matching the feature amount of the character pattern obtained in the feature amount acquisition step with the feature amount of the dictionary template of each character, and classifying the character pattern,
The feature synthesis step includes
For each scanning direction, the vertical coordinate contained in the division position information in the direction perpendicular to the scanning direction obtained in the division position information acquisition step and the scanning direction obtained in the outline feature extraction step A step of obtaining the edge direction and edge strength at each contour point of the coordinate line of the contour feature in each coordinate from the edge feature obtained in the edge feature extraction step based on the contour feature of
Obtaining a feature storage position on the memory determined by the degree information and direction of the outline point and the division position included in the division position information;
Adding a weight in a divided frame corresponding to a vertical coordinate included in the divided position information to the edge strength for each edge direction obtained in the step, and accumulating;
A character recognition method comprising: A character reading unit that obtains binarized image data obtained by reading a document including a character to be recognized, and a character cutout unit that cuts out a binarized character pattern from the binarized image data in units of one character. A feature extraction unit that extracts features from a binarized character pattern, and a character recognition unit that obtains a recognition result by comparing the feature amount extracted by the feature extraction unit with a plurality of dictionary templates stored in the character recognition dictionary A character recognition method in a character recognition device comprising:
The feature extraction unit includes:
Rectangle information relating to a rectangle circumscribing the binarized character pattern in the ordinate / horizontal coordinate system and rectangle information relating to the rectangle circumscribing in the oblique coordinate system obtained by rotating the ordinate / horizontal coordinate system by 45 ° to the character pattern are extracted. Rectangle information extraction step to be stored in the memory,
An outline feature extraction step for sequentially extracting outline points of the character pattern for each scanning direction and storing them in a memory as primary features, secondary features,..., N-th feature;
An edge feature extraction step of extracting an edge feature of the character pattern and storing it in a memory;
For the direction perpendicular to the scanning direction in the outer feature extraction step based on the edge position obtained from the extracted edge feature at the edge feature extraction step, in the vertical direction while scanning a pre Symbol character pattern edge between density extracting step of storing in the memory by extracting an edge between density from the scanning direction or falling edge of di period of,
Based on the rectangular information extracted in the rectangular information extracting step and the inter-edge density extracted in the inter-edge density extracting step, a rectangle circumscribing the character pattern in each scanning direction and in a direction perpendicular to each scanning direction is vertically The division frame to be divided is determined, the division position of each coordinate in which the vertical coordinate in the rectangle exists, the division position of each coordinate is obtained, and the vertical coordinate line in the rectangle is assigned to which division frame obtains the division position of existing or respective coordinate lines, stores obtain a weight indicating the ratio of the edge between the density occupying a split position and coordinate lines dividing position of each coordinate line for each direction coordinates in the rectangle in the memory Division position information acquisition step,
A feature synthesis step for performing feature synthesis for each of the scanning directions;
An inclination detecting step for detecting an inclination of the character pattern after the feature synthesis is completed;
In accordance with the inclination detected in the inclination detection step, a feature amount acquisition step for obtaining a feature amount of the character pattern normalized based on the accumulated edge strength,
With
The character recognizing unit includes a character recognition step of matching the feature amount of the character pattern obtained in the feature amount acquisition step with the feature amount of the dictionary template of each character, and classifying the character pattern,
The feature synthesis step includes
For each scanning direction, the vertical coordinate contained in the division position information in the direction perpendicular to the scanning direction obtained in the division position information acquisition step and the scanning direction obtained in the outline feature extraction step A step of obtaining the edge direction and edge strength at each contour point of the coordinate line of the contour feature in each coordinate from the edge feature obtained in the edge feature extraction step based on the contour feature of
Obtaining a feature storage position on the memory determined by the degree information and direction of the outline point and the division position included in the division position information;
Adding a weight in a divided frame corresponding to a vertical coordinate included in the divided position information to the edge strength for each edge direction obtained in the step, and accumulating;
Consist of,
A character recognition method characterized by the above.   In the rectangular information extracting step, after extracting rectangular information related to a rectangle circumscribing the character pattern in the vertical and horizontal coordinate system, two diagonal lines of the rectangle are obtained, and for each diagonal line, an oblique coordinate system is set in the diagonal direction. The character recognition method according to claim 1, further comprising a step of setting as a coordinate system and performing feature extraction in the diagonal direction of the rectangle as the feature extraction in the oblique direction.   In the division position information acquisition step, when determining a division frame for dividing the rectangle of the character pattern in an oblique direction, the rectangle of the character pattern obtained by extracting the central division line in the vertical and horizontal coordinate system is divided into two equal parts. The character recognition method according to claim 1, further comprising a step of fixing at a center.   The character recognizing unit is centered on a diagonal position of a rectangle of the character pattern extracted in a vertical and horizontal coordinate system and a division position information of a feature storage position of the diagonal feature amount with respect to a matching result of the diagonal feature amount. The character recognition method according to claim 1, further comprising a step of weighting by applying a weight determined by a weight function that is symmetric with respect to an oblique 45 ° line.

Images