JP4330296B2

JP4330296B2 - Character recognition device, character recognition reliability determination method, execution program thereof, and recording medium storing the same

Info

Publication number: JP4330296B2
Application number: JP2001282429A
Authority: JP
Inventors: 明中村; 博光川尻
Original assignee: Sanyo Electric Co Ltd
Current assignee: Sanyo Electric Co Ltd
Priority date: 2001-09-17
Filing date: 2001-09-17
Publication date: 2009-09-16
Anticipated expiration: 2021-09-17
Also published as: JP2003091699A

Description

【０００１】
【発明の属する技術分野】
本発明は、入力情報を処理して所定の文字を認識する文字認識装置および対象候補文字の信頼度を判定する文字認識信頼度判定方法に関する。
【０００２】
【従来の技術】
従来の文字認識方法においては、たとえば筆記入力された文字の特徴量を抽出し、これを認識辞書中の特徴量と比較して、両者の類似度が高い、もしくは両者の距離値が小さい（これらをまとめて、便宜上、「確信度」が高いと称する）認識候補文字を出力するようにしていた。しかしながら、かかる一文字毎の文字認識では、筆記文字が認識辞書の特徴量に近接している場合には比較的精度の良い認識結果が得られるが、認識辞書の特徴量から離れた文字を筆記した場合には、適正な認識結果を簡単に得ることができない。
【０００３】
そこで、かかる一文字毎の文字認識に加え、前後の文字あるいは単語間・文節間の連接確率ないし共起確率を検出し、上記文字毎の確信度とこれらの確率とから文字列の整合度を算出し、かかる整合度に従って文字列全体の認識文字列候補を出力する、いわゆる後処理が実行されている。ところが、かかる後処理の際に、あまりに多くの認識候補文字を対象とすると、後処理の計算処理時間が増大してしまう。また、確信度の低い認識候補文字を対象とすると、後処理の結果、却って誤った文字列候補を出力する恐れもある。
【０００４】
そこで、確信度の低い認識候補文字を後処理の対象から除外し、これにより後処理の対象文字数を効果的に低減させる方法が採られる。たとえば、それぞれの認識候補文字とその直前および直後の確信度第１位の文字との間の連接確率をしきい値と比較し、これがしきい値に満たない場合に、当該認識候補文字を「誤」と判別して後処理の対象から外すようにする。すなわち、図９において、認識候補文字群の内、正誤判定対象の文字が「ボ」であるとすると、その直前の第１位の認識候補文字「と」との間の連接確率Ｐ１と、その直後の第１位の認識候補文字「不」との間の連接確率Ｐ２を、それぞれ連接確率辞書から求め、これらＰ１とＰ２を乗じたＰ３がしきい値を超えているか否かで、当該判定文字「ボ」の正誤を判別する。同様に、「ボ」以外の認識候補文字と、その直前の第１位の認識候補文字「と」およびその直後の第１位の認識候補文字「不」との間の連接確率Ｐ１、Ｐ２からＰ３を求め、これがしきい値を超えるか否かで、当該認識候補文字の正誤を判定する。そして、Ｐ３がしきい値を超えない認識候補文字を後処理の対象から除外し、これにより後処理の効率化を図るのである。
【０００５】
【発明が解決しようとする課題】
しかしながら、かかる方法によれば、連接確率としきい値とを単純に比較するものであるから、正しい認識結果（正読）だが連接確率の低い場合や、逆に、誤った認識結果（誤読）だが連接確率の高い場合に、正誤の判定が不適切なものとなってしまう可能性が高い。
【０００６】
また、直前と直後の第１位の認識候補文字が誤りの場合には、判定対象文字の正誤判定を正確に行うことができなくなる。
【０００７】
したがって、上記従来技術によれば、本来後処理の対象から除外すべき認識候補文字を信頼度の高いものとして後処理の対象として出力し、逆に、本来後処理の対象とすべき認識候補文字を信頼度の低いものとして後処理の対象から除外する結果が生じ、信頼度の判定によって却って認識候補文字の精度を低下させる結果ともなっていた。
【０００８】
そこで本発明は、上記問題を解消し、正読および誤読に対する連接確率の設定が不適当な場合や、直前と直後の第１位の認識候補文字が誤りである場合にも、比較的精度よく認識候補文字の信頼度を判定でき、もって、認識候補文字の精度を向上させ得る文字認識装置および文字認識信頼度判定方法を提供することを課題とするものである。
【０００９】
【課題を解決するための手段】
本発明は、判定対象の候補文字群とその直前および／もしくは直後の候補文字群との間の連接確率の値を当該判定対象の候補文字群の特徴量とし、かかる特徴量をサンプルデータもしくはサンプルデータの統計的傾向と比較することにより当該判定対象の候補文字群の信頼度を判定するものである。
【００１０】
請求項１の発明は、文字認識装置に関するものであって、入力情報を処理して認識候補文字群を出力する文字認識手段と、前記文字認識手段によって出力される一の入力情報に対する判定対象認識候補文字群中の各認識候補文字とその直前の入力情報に対する直前認識候補文字群との間の連接確率の値および／もしくはその直後の入力情報に対する直後認識候補文字群との間の連接確率の値を前記判定対象認識候補文字群の特徴量として抽出する特徴抽出手段と、前記特徴抽出手段からの特徴量からなる連接特徴ベクトルと前記判定対象認識候補文字群の第１位の認識候補文字が正読または誤読であることが既知である予め抽出された連接特徴ベクトルであるサンプルデータの正読、誤読のそれぞれの集合の分布に基づいて前記判定対象認識候補文字群の信頼度を算出する信頼度算出手段と、前記信頼度算出手段からの信頼度に基づいて前記判定対象認識候補文字群から後処理の対象として出力する認識候補数を設定する認識候補数制御部とを有することを特徴とする。
さらに、前記特徴抽出手段は、前記判定対象認識候補文字群中の一の認識候補文字とその直前または直後認識候補文字群中の１位からＭ位までの各認識候補文字との間の連接確率の内、最高の連接確率を当該一の認識候補文字と前記直前または直後認識候補文字群との間の連接確率とすることを特徴とする。
【００１１】
請求項２の発明は、請求項１において、前記特徴抽出手段は、前記判定対象認識候補文字群中の各認識候補文字と前記直前認識候補文字群中の最上位確信度の認識候補文字との間の連接確率の値および／もしくは前記直後認識候補文字群中の最上位確信度の認識候補文字との間の連接確率の値を当該判定対象認識候補文字群の特徴量として抽出することを特徴とする。
【００１３】
請求項３の発明は、請求項１または２において、前記特徴抽出手段は、前記連接確率の値と、前記判定対象認識候補文字群中の各認識候補文字の確信度の値とによって当該判定対象認識候補文字群の特徴量を抽出することを特徴とする。
【００１４】
請求項４の発明は、請求項１ないし３の何れかにおいて、前記特徴抽出手段は、前記直前認識候補文字群が存在しない場合には前記判定対象認識候補文字群と直後認識候補文字群との間の連接確率を前記直前認識候補文字群との間の連接確率として代用し、前記直後認識候補文字群が存在しない場合には前記判定対象認識候補文字群と直前認識候補文字群との間の連接確率を前記直後認識候補文字群との間の連接確率として代用することを特徴とする。
【００１５】
請求項５の発明は、請求項１ないし４の何れかにおいて、前記信頼度算出手段は、前記特徴量から前記判定対象認識候補文字群中の一の認識候補文字の確からしさを判別得点として算出する判別得点算出手段を含み、前記当該判別得点に基づいて前記信頼度を算出することを特徴とする。
【００１６】
請求項６の発明は、請求項１ないし５の何れかにおいて、前記後処理制御手段は、前記信頼度算出手段から算出された信頼度に基づいて、後処理の対象とする認識候補文字を制限することを特徴とする。
【００１７】
請求項７の発明は、一連の候補文字の連接関係によって一の候補文字の信頼度を判定する文字認識信頼度判定方法であって、判定対象候補文字群中の各候補文字と、その直前の候補文字群との間の連接確率の値および／もしくはその直後の候補文字群との間の連接確率の値を前記判定対象候補文字群の特徴量として抽出する特徴抽出ステップと、前記特徴抽出ステップからの特徴量からなる連接特徴ベクトルと前記判定対象候補文字群の第１位の認識候補文字が正読または誤読であることが既知である予め抽出された連接特徴ベクトルであるサンプルデータの正読、誤読のそれぞれの集合の分布に基づいて前記判定対象候補文字群の信頼度を算出する信頼度算出ステップとを有することを特徴とする。
さらに、前記特徴抽出ステップは、前記判定対象候補文字群中の一の候補文字とその直前または直後の候補文字群中の１位からＭ位までの各候補文字との間の連接確率の内、最高の連接確率を当該一の候補文字と前記直前または直後の候補文字群との間の連接確率とすることを特徴とする。
【００１８】
請求項８の発明は、請求項７において、前記特徴抽出ステップは、前記判定対象候補文字群中の各候補文字と前記直前の候補文字群中の最上位確信度の候補文字との間の連接確率の値および／もしくは前記直後の候補文字群中の最上位確信度の候補文字との間の連接確率の値を当該判定対象候補文字群の特徴量として抽出することを特徴とする。
【００２０】
請求項９の発明は、請求項７または８において、前記特徴抽出ステップは、前記連接確率の値と、前記判定対象候補文字群中の各候補文字の確信度の値とによって当該判定対象候補文字群の特徴量を抽出することを特徴とする。
【００２１】
請求項１０の発明は、請求項７ないし９の何れかにおいて、前記特徴抽出ステップは、前記直前の候補文字群が存在しない場合には、前記判定対象候補文字群と前記直後の候補文字群との間の連接確率を前記直前の候補文字群との間の連接確率として代用し、前記直後の候補文字群が存在しない場合には、前記判定対象候補文字群と前記直前の候補文字群との間の連接確率を前記直後の候補文字群との間の連接確率として代用することを特徴とする。
【００２２】
請求項１１の発明は、請求項７ないし１０の何れかにおいて、前記信頼度算出ステップは、前記特徴量から前記判定対象候補文字群中の一の候補文字の確からしさを判別得点として算出する判別得点算出ステップを含み、当該判別得点に基づいて前記信頼度を算出することを特徴とする。
【００２３】
請求項１２の発明は、請求項７ないし１１の何れかの文字認識信頼度判定方法における各処理ステップを実行するプログラムである。
【００２４】
請求項１３の発明は、請求項７ないし１１の何れかの文字認識信頼度判定方法における各処理ステップを実行するプログラムを記憶した記録媒体である。
【００２５】
本発明の特徴は、以下に示す実施の形態の説明により更に明らかとなろう。
【００２６】
なお、請求項における「文字認識手段」は実施の形態における文字認識部２および文字認識辞書３が対応する。請求項における「特徴抽出手段」は実施の形態における連接特徴抽出部４および文字間連接確率辞書５が対応する。請求項における「信頼度算出手段」は実施の形態における判別得点算出部６、認識信頼度算出部７および判別得点―信頼度変換テーブル８が対応する。請求項における後処理制御手段」は実施の形態における認識候補数制御部９および信頼度―累積正読率テーブル１０が対応する。
【００２７】
また、請求項における「特徴抽出ステップ」は実施の形態における図７のステップＳ１０３が対応する。請求項における「信頼度算出ステップ」は実施の形態における図７のステップＳ１０４およびＳ１０５が対応する。
【００２８】
ただし、以下の実施の形態は、あくまでも、本発明の一つの実施形態であって、本発明ないし各構成要件の用語の意義は、以下の実施の形態に記載されたものに制限されるものではない。
【００２９】
【発明の実施の形態】
以下、本発明の実施の形態につき図面を参照して説明する。
【００３０】
まず、図１に実施の形態に係る手書き文字認識装置の回路ブロック図を示す。図において、１は入力部で、タブレット等に手書き入力された筆跡から筆跡文字情報を生成し出力する。２は文字認識部で、入力部１から供給された筆跡文字情報を文字認識辞書３の文字特徴量と比較し、両者の近接度（確信度）が１位からＮ位までの認識辞書中の文字を当該筆跡文字の認識候補文字として出力する。３は文字認識辞書で、候補文字がその文字特徴量とともに記憶されている。
【００３１】
４は連接特徴抽出部で、文字認識部２から供給される認識候補文字列の内、正誤判定対象である一の認識候補文字の連接特徴量を、文字間連接確率辞書５を参照しながら抽出する。５は文字間連接辞書で、所定の２つの文字が連接する確率を当該２つの文字に対応づけて記憶している。
【００３２】
６は判別得点算出部で、連接特徴抽出部４からの連接特徴量を処理して、当該判定対象の認識候補文字群の正誤判別得点を算出する。
【００３３】
７は認識信頼度算出部で、判別得点算出部６からの判別得点と、判別得点−信頼度変換テーブル８とを比較して、当該判定対象の認識候補文字群の信頼度を出力する。８は判別得点−信頼度変換テーブルで、判別得点と信頼度の関係をテーブルとして記憶するものである。
【００３４】
９は認識候補数制御部で、認識信頼度７からの信頼度と、信頼度−累積正読率テーブル１０とを比較して、当該判定対象文字の認識候補数を制限するものである。１０は信頼度−累積正読率テーブルで、信頼度と正読率の関係をテーブルとして記憶するものである。
【００３５】
１１は後処理部で、認識候補数制御部９によって設定された個数の認識候補文字を対象として後処理を行う。例えば、図９の例において、「こ」の筆記文字に対する認識候補数がＭ１個、以下、「と」、「ボ」、「不」、「可」、「欠」、「で」の筆記文字に対する認識候補数がＭ２、Ｍ３、Ｍ４、Ｍ５、Ｍ６、Ｍ７個と認識候補数制御手段９によって設定されたとすると、文字認識部２で生成されたこれら各文字の認識候補文字の内、それぞれ確信度が上位Ｍ１、Ｍ２、Ｍ３、Ｍ４、Ｍ５、Ｍ６、Ｍ７個の認識候補文字のみを対象として、後処理部１１にて、後処理が行われる。
【００３６】
次に、上記回路ブロック図に示された各部の処理の詳細について説明する。まず、図２を参照して、連接特徴抽出部４の処理について説明する。かかる連接特徴抽出部４では、判定対象の認識候補文字群とその直前の第１位候補文字との間の連接確率Ｐｂｋと、その直後の第１位候補文字との間の連接確率Ｐｆｋの何れか一方または両方を文字間連接確率辞書５から読み出し、当該認識候補文字群の連接特徴量とする。
【００３７】
例えば、連接確率ＰｂｋとＰｆｋの両方を用いて連接特徴量を抽出する場合には、図２に示すように、判定対象文字の認識候補文字群（確信度１位からＮ位まで）である「目」、「日」、…、「田」と、その直前の認識候補文字群の確信度１位文字「今」との間の連接確率Ｐｂ１、Ｐｂ２、…、ＰｂＮと、その直後の確信度１位文字「は」との間の連接確率Ｐｆ１、Ｐｆ２、…、ＰｆＮをそれぞれ文字間連接確率辞書５から読み出し、これら連接確率の値を、当該判定対象の認識候補文字群の連接特徴量とする。
【００３８】
なお、上記では、認識候補文字群の連接特徴量をＰｂｋとＰｆｋの両方にて抽出するようにしたが、ＰｂｋまたはＰｆｋの何れか一方にて抽出するようにしても良い。また、ＰｂｋとＰｆｋの両方にて抽出する場合において、直前の第１位認識候補文字が存在しない場合（例えば、判定対象文字が文字列の先頭であるような場合）には、当該候補文字の直後文字に対する連接確率Ｐｆｋをそのまま直前文字に対する連接確率Ｐｂｋに代用して連接特徴量を抽出する。同様に、直後の第１位認識候補文字が存在しない場合（例えば、判定対象文字が文字列の末尾であるような場合）には、Ｐｂｋをそのまま代用する。
【００３９】
次に、図３を参照して、判別得点算出部６における処理について説明する。上記連接特徴抽出部４にて抽出したＰｂｋとＰｆｋの値は、２Ｎ次元のベクトル空間において所定のベクトル（連接特徴ベクトル）として表現できる。判別得点算出部６では、予め第１位の認識候補文字が正読または誤読であるサンプルについて連接特徴ベクトルを学習データとして抽出しておき、これと判定対象の認識候補文字群の連接特徴ベクトルとを比較して、当該認識候補文字群の判別得点を算出する。
【００４０】
たとえば、正読と誤読の連接特徴ベクトルの学習データと判定対象の認識候補文字群の連接特徴ベクトルが図３に示すような状態にあるとする。判別得点算出部６は、正読、誤読のそれぞれの集合（クラス）の特徴ベクトルの分布から、予め、両クラスの特徴ベクトルの平均値（重心ベクトル）および共分散行列を求め、これを記憶している。そして、これら各クラスの重心と判定対象文字群の連接特徴ベクトルとの間のマハラノビス距離ＤＭｃ、ＤＭｅを求め、これらの値の比、比の対数または差を判別得点とする。
【００４１】
ここで、上記マハラノビス距離ＤＭは次のようにして算出される。すなわち、クラスＣ_１の重心ベクトルをｍ_１、クラスＣ_１の共分散行列をΣ_１とすると、所定の特徴ベクトルｘからｍ_１へのマハラノビス２乗距離ＤＭ_１は、次式で定義される。
【００４２】
ＤＭ_１＝（ｘ−ｍ_１）^tΣ１^-1（ｘ−ｍ_１）（ｔ：転置、−１：逆行列）ここで、共分散行列Σ_１はｎ×ｎの正方行列であり（ｎ：特徴空間の次元数）、その（ｉ,ｊ）要素はｉ番目の特徴量とｊ番目の特徴量の共分散、すなわちΣ_１（ｉ,ｊ）＝σ_ｉｊである。
【００４３】
なお、上記では、判別得点をマハラノビス距離ＤＭを用いて算出したが、これに替えて、正読、誤読の各クラスの特徴ベクトルの分布から線形判別分析により線形判別関数を求めておき、判定対象の認識候補文字群の特徴ベクトルに対してこの線形判別関数を当てはめて判別得点を求めるようにしても良い。また、正読、誤読の学習サンプルから抽出した特徴ベクトルを学習データとして、対象の特徴ベクトルが正読か誤読かを判定できるように学習させたニューラルネットを用い、判定対象の特徴ベクトルに対する当該ニューラルネットの出力値を判別得点とするようにしてもよい。
【００４４】
次に、図４を参照して認識信頼度算出部７の処理について説明する。たとえば、上記判別得点の算出において、マハラノビス距離ＤＭｃ、ＤＭｅの距離の比または比の対数を判別得点とした場合、正読および誤読の各学習サンプルから得られる判別得点と信頼度の関係は図４に示すようになる。ここで、信頼度は、学習サンプルからベイズの定理によって算出される。
【００４５】
すなわち、判別得点ｙを有する正読サンプル個数の全正読サンプル個数に対する比率をｐ（ｙ｜Ｘ１＝Ｃ）、判別得点ｙを有する誤読サンプル個数の全誤読サンプル個数に対する比率をｐ（ｙ｜Ｘ１＝Ｅ）、全サンプル数に対する正読サンプル数の総数の比率をＰ（Ｘ１＝Ｃ）、全サンプル数に対する誤読サンプル数の総数の比率をＰ（Ｘ１＝Ｅ）とすると、判別得点ｙを有する認識候補文字群の確信度１位の認識候補文字の信頼度は、次式によって算出できる。
【００４６】
Ｐ（Ｘ１＝Ｃ｜ｙ）＝ｐ（ｙ｜Ｘ１＝Ｃ）・Ｐ（Ｘ１＝Ｃ）／[ｐ（ｙ｜Ｘ１＝Ｃ）・Ｐ（Ｘ１＝Ｃ）＋ｐ（ｙ｜Ｘ１＝Ｅ）・Ｐ（Ｘ１＝Ｅ）]
ここで、Ｘ１は１位の認識候補文字を表し、Ｘ１＝Ｃ、Ｘ１＝Ｅはそれぞれ、１位認識候補文字が正解、不正解である事象を意味する。
【００４７】
かかる式から判別得点と信頼度の関係を示す判別得点―信頼度変換テーブルを予め作成しておき、これを判別得点―信頼度変換テーブル８に記憶させておく。認識信頼度算出部７は判別得点算出部６からの判別得点と、当該判別得点―信頼度変換テーブル８の得点を比較し、該当する信頼度を、当該認識候補文字群の第１位の認識候補文字の信頼度として出力する。
【００４８】
次に、図５を参照して、認識候補数制御部９の処理について説明する。図５の上部に示す表は、判定対象の確信度１位の認識候補文字に対する信頼度と、当該判定対象のＮ位までの認識候補文字の中に正読の文字が含まれる累積確率との関係を示すものである。かかる表中の確率は、上記正読、誤読の学習サンプルを基に予め算出しておく。
【００４９】
信頼度―累積正読率テーブル１０には、かかる表を記憶させておく。そして、認識候補数制御部９は、判定対象の認識候補文字群の信頼度と当該テーブル中の信頼度レベルとを比較し、該当する信頼度レベルの累積確率を参照しながら何位までの認識候補文字を言語処理部１１に出力するかを決定する。ここで、何位までを出力するかは、例えば、該当する信頼度レベルの累積確率が所定のしきい値に達したか否かで決定する。この際、設定されるしきい値は、全ての信頼度レベルに対して一律としても良いし、あるいは、信頼度レベル毎に個別に設定するようにしても良い。
【００５０】
あるいは、図５の上部の表を基に、信頼度レベル毎の出力候補数を予め設定し、これを信頼度−累積正読率テーブル１０に記憶させておいても良い。図５の下部に示す表は、信頼度レベルと出力候補数とを予め設定した場合の一例である。信頼度―累積正読率テーブル１０に予めかかる表を記憶させた場合には、認識候補数制御部９は、該当する出力候補数を表から読み出し、それに従って、言語処理部１１に出力される認識候補文字を制限する。
【００５１】
以上の実施の形態においては、認識候補文字毎に個別に連接確率を比較するのではなく、認識候補文字群の連接確率の値の組み合わせの統計的傾向に基づいて認識結果の信頼度を算出し、これにより認識候補出力個数を設定するものであるから、正読だが連接確率の低い連接関係や、誤読だが連接確率の高い連接関係が個別に存在するとしても、かかる連接関係が信頼度判定に影響するのを抑制でき、もって、正読率の高い認識候補文字を言語処理部に出力することができるようになる。
【００５２】
次に、本発明に係る第２の実施形態について、以下に説明する。本実施形態においては、上記連接特徴抽出部２における連接特徴量の抽出処理を変更するものである。すなわち、上記実施の形態では、判定対象の認識候補文字の連接確率（Ｐｂｋ、Ｐｆｋ）として、当該判定対象の認識候補文字と直前の第１位文字または直後の第１位文字との連接確率を採用したが、本実施の形態では、判定対象の認識候補文字と直前の１位からＭ位までの文字との間の連接確率の最大値をＰｂｋとし、同様に、判定対象の認識候補文字と直前の１位からＭ位までの文字との間の連接確率の最大値をＰｆｋとする。
【００５３】
たとえば図６の例においては、判定対象の認識候補文字の１位文字「日」に対するＰｂ１は、当該「日」と直前の１位文字「朋」からＭ位文字「胡」までのそれぞれの連接確率Ｐ（Ｃ１｜Ｃｂｋ）の内、最大の連接確率を採用する。また、１位文字「日」に対するＰｆ１は、当該「日」と直後の１位文字「も」からＭ位文字「亡」までのそれぞれの連接確率Ｐ（Ｃｆｋ｜Ｃ１）の内、最大の連接確率を採用する。同様に、判定対象の認識候補文字の２位文字「月」に対するＰｂ２、Ｐｆ２は、直前、直後の文字群に対する連接確率の最大値をそれぞれ採用する。
【００５４】
ここで、Ｃ１は判定対象の認識候補１位の文字を表し、Ｃｂｋ、Ｃｆｋはそれぞれ、直前、直後の認識候補ｋ位の文字を表す。そして、Ｐ（Ｃｊ｜Ｃｉ）は、文字Ｃｉに続いて文字Ｃｊが現れる連接確率を表す。
【００５５】
かかる第２の実施形態においては、直前・直後の文字群との間の最大の連接確率をＰｂｋ、Ｐｆｋとして採用するものであるから、上記第１の実施形態における作用効果に加え、さらに、直前・直後の第１位の認識候補文字が誤りである場合にも、高精度の信頼度判定を行えるものである。
【００５６】
さらに他の実施形態として、上記連接確率Ｐｂｋ、Ｐｆｋの他、第Ｌ位までの認識候補文字の確信度（類似度もしくは距離値）を特徴要素として加え、（２Ｎ＋Ｌ）次元のベクトル空間にて当該認識候補文字群の特徴ベクトルを抽出するようにしてもよい。かかる場合には、図３に示す判別空間は（２Ｎ＋Ｌ）次元となる。また、正読・誤読のサンプルも、当該サンプルの認識候補文字群に対する連接確率Ｐｂｋ、Ｐｆｋの他、Ｌ位までの当該認識候補文字の確信度が特徴量抽出要素とされ、かかるサンプルデータに従って判別得点―信頼度変換テーブル８と信頼度―累積正読率テーブル１０に記憶されるテーブルが設定される。かかる第３の実施の形態では、連接関係のみならず確信度が加味されるものであるから、より高精度の信頼度判定が可能となる。
【００５７】
ところで、上記実施の形態では、図１におけるブロック毎に処理を分けて一連の処理フローを説明したが、制御プログラムに従ってＣＰＵによってかかる処理フローを実行することも可能である。かかる場合、上記処理フローは、ＲＯＭまたはＲＡＭに制御プログラムとして記憶される。また、文字認識辞書３、文字間連接確率辞書５、判別得点―信頼度変換テーブル８および信頼度―累積正読率テーブル１０の参照データもＲＯＭまたはＲＡＭに記憶される。ＣＰＵは、かかる制御プログラムに従って、参照データを参照しながら、上記の処理を実行する。
【００５８】
図７に、かかる制御プログラムによるフローを示す。ここで、ステップＳ１０１は上記入力部１における処理、ステップＳ１０２は上記文字認識部２における処理、ステップＳ１０３は上記連接特徴抽出部４における処理、ステップＳ１０４は上記判別得点算出部６における処理、ステップＳ１０５は上記認識信頼度算出部７における処理、ステップＳ１０６は上記認識候補数制御部９における処理である。
【００５９】
かかる制御プログラムおよび各種参照データは、フロッピーディスク等の記録媒体またはインターネット等の伝送媒体を介して取引され得る。記録媒体または伝送媒体を介して取引されるデータのファイル構造の一例を図８に示す。記録媒体には、かかるファイル構造のデータが記録される。また、伝送媒体を介した取引では、かかるファイル構造のデータが伝送媒体を介して供給される。
【００６０】
ただし、判定対象の候補文字の信頼度を判定するのみであれば、図７に示す全てのステップに関するデータを供給する必要はなく、この内、ステップＳ１０３ないしＳ１０５のステップの実行プログラムおよびそれに必要な文字間連接確率辞書データ、判別得点―信頼度変換テーブルデータのみを供給すれば良い。かかる場合、供給される記録媒体にはこれらのデータがファイルデータとして記憶され、また、伝送媒体を介して供給する場合には、かかるファイルデータが伝送される。
【００６１】
以上、本発明に係る実施の形態について説明したが、本発明は係る実施の形態に制限されるものではなく、他に種々の変更が可能である。たとえば、上記実施の形態では、本発明に係る信頼度判定方法を手書き文字認識装置に実施した例を示したが、本発明にかかる信頼度判定方法はこれに限定されるものではなく、光学式文字認識装置あるいは音声入力による認識装置等にも広く採用し得るものである。
【００６２】
本発明に係る信頼度判定方法の特徴は、上記実施の形態でいえば、図７のステップＳ１０３ないしＳ１０５にあり、直前と直後の文字の両方あるいは何れか一方の文字との間の連接関係の分布状態によって判別得点を算出し、かかる判別得点に基づいて対象文字の信頼度を判定するところにある。
【００６３】
したがって、かかるステップＳ１０３ないしＳ１０５の処理は、上記実施形態に係る手書き文字認識装置のみならず、これ以外の認識装置においても用いることができる。例えば、図７のステップＳ１０１とＳ１０２を周知の光学式文字認識装置あるいは音声入力による認識処理としてよい。また、ステップＳ１０６も認識候補数の制限のみならず、信頼度に応じて、当該認識結果をリジェクト（無効）するようにしても良い。
【００６４】
その他、ステップＳ１０４における判別得点の算出方法や、ステップＳ１０５における認識信頼度の算出方法も、上記実施の形態にて示したマハラノビス距離ＤＭを用いる方法や、ベイズの定理を用いる方法以外の方法を採用することもできる。
【００６５】
本発明の実施形態は、本発明の技術的思想の範囲内において、適宜、様々な変更が可能である。
【００６６】
【発明の効果】
以上、本発明によれば、正読および誤読に対する連接確率が不適当な場合や、直前と直後の確信度１位の候補文字が誤りである場合にも、比較的精度よく判定対象文字の信頼性を判定でき、もって、これを文字認識装置に採用した場合には、認識候補文字の精度を向上させることができるようになる。
【図面の簡単な説明】
【図１】実施の形態に係る回路ブロック図を示す図
【図２】実施の形態に係る連接特徴抽出部の処理を説明するための図
【図３】実施の形態に係る判別得点算出部の処理を説明するための図
【図４】実施の形態に係る認識信頼度算出部の処理を説明するための図
【図５】実施の形態に係る認識候補数制御部の処理を説明するための図
【図６】実施の形態に係る連接特徴抽出部の他の処理を説明するための図
【図７】実施の形態に係る実行フローチャート
【図８】実施の形態に係る実行プログラムと参照データのファイル構造
【図９】従来例を説明するための図
【符号の説明】
１入力部
２文字認識部
３文字認識辞書
４連接特徴抽出部
５文字間連接確率辞書
６判別得点算出部
７認識信頼度算出部
８判別得点―信頼度変換テーブル
９認識候補数制御部
１０信頼度−累積正読率テーブル[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a character recognition device that processes input information to recognize a predetermined character and a character recognition reliability determination method that determines the reliability of a target candidate character.
[0002]
[Prior art]
In the conventional character recognition method, for example, the feature amount of a character input by writing is extracted, and compared with the feature amount in the recognition dictionary, the similarity between the two is high, or the distance value between the two is small (these For the sake of convenience, the recognition candidate character is output (referred to as “high confidence”). However, in such character recognition for each character, when the written character is close to the feature value of the recognition dictionary, a relatively accurate recognition result can be obtained, but the character that is far from the feature value of the recognition dictionary is written. In this case, a proper recognition result cannot be easily obtained.
[0003]
Therefore, in addition to character recognition for each character, the connection probability or co-occurrence probability of the preceding and following characters or between words and phrases is detected, and the consistency of the character string is calculated from the certainty factor for each character and these probabilities. Then, so-called post-processing is performed in which recognized character string candidates for the entire character string are output in accordance with the degree of matching. However, if too many recognition candidate characters are targeted during such post-processing, the post-processing calculation processing time increases. In addition, if a recognition candidate character with a low certainty factor is targeted, an erroneous character string candidate may be output instead as a result of post-processing.
[0004]
Therefore, a method is adopted in which recognition candidate characters with low certainty are excluded from the targets for post-processing, thereby effectively reducing the number of target characters for post-processing. For example, the probability of connection between each recognition candidate character and the character with the first degree of certainty immediately before and immediately after that is compared with a threshold value. It is discriminated as “false” and is excluded from the post-processing target. That is, in FIG. 9, if the character to be judged is “Bo” in the recognition candidate character group, the connection probability P1 between the first recognition candidate character “and” immediately before that, The connection probability P2 between the first recognition candidate character “not” immediately after is obtained from the connection probability dictionary, and whether or not P3 multiplied by P1 and P2 exceeds a threshold value The correctness of the character “B” is determined. Similarly, from the connection probabilities P1 and P2 between the recognition candidate character other than “BO”, the first recognition candidate character “to” immediately before it, and the first recognition candidate character “not” immediately after that. P3 is obtained, and whether or not the recognition candidate character is correct is determined based on whether or not this exceeds a threshold value. Then, the recognition candidate characters whose P3 does not exceed the threshold are excluded from the targets for post-processing, thereby improving the efficiency of the post-processing.
[0005]
[Problems to be solved by the invention]
However, according to such a method, since the connection probability is simply compared with the threshold value, the correct recognition result (correct reading), but the connection probability is low, or conversely, the incorrect recognition result (misreading). When the connection probability is high, there is a high possibility that the correctness / incorrectness determination is inappropriate.
[0006]
In addition, when the first recognition candidate character immediately before and immediately after is incorrect, the correctness determination of the determination target character cannot be performed accurately.
[0007]
Therefore, according to the above prior art, recognition candidate characters that should be excluded from the post-processing target are output as post-processing targets with high reliability, and conversely, recognition candidate characters that should originally be the target of post-processing. As a result, the result of excluding the character from the post-processing target is generated and the accuracy of the recognition candidate character is lowered by the determination of the reliability.
[0008]
Therefore, the present invention solves the above-mentioned problem, and relatively accurately when the connection probability for correct reading and misreading is inappropriate or when the first recognition candidate character immediately before and immediately after is incorrect. It is an object of the present invention to provide a character recognition device and a character recognition reliability determination method that can determine the reliability of recognition candidate characters and thereby improve the accuracy of recognition candidate characters.
[0009]
[Means for Solving the Problems]
In the present invention, the value of the connection probability between the candidate character group to be determined and the immediately preceding and / or immediately following candidate character group is used as the feature amount of the candidate character group to be determined, and the feature amount is sampled data or sample The reliability of the candidate character group to be determined is determined by comparing with the statistical tendency of the data.
[0010]
The invention according to claim 1 relates to a character recognition device, comprising: character recognition means for processing input information and outputting a recognition candidate character group; and recognition object recognition for one input information output by the character recognition means. The value of the connection probability between each recognition candidate character in the candidate character group and the immediately preceding recognition candidate character group for the input information immediately before and / or the connection probability between the immediately following recognition candidate character group for the input information immediately after that A feature extraction unit that extracts a value as a feature amount of the determination target recognition candidate character group, a concatenated feature vector including the feature amount from the feature extraction unit, and a first recognition candidate character of the determination target recognition candidate character group. sample data right reading is positive read a concatenated feature vector extracted in advance are known to be misread, the determination target recognition based on the distribution of each set of misreading Recognition candidates to set the reliability calculation means calculates the reliability of the auxiliary character group, the recognition candidate number outputted as a target of the post-processing from the determination target recognition candidate character group based on the reliability from the reliability calculation unit And a number controller .
Further, the feature extraction means may determine a connection probability between one recognition candidate character in the determination target recognition candidate character group and each recognition candidate character from the first position to the M position in the immediately preceding or immediately following recognition candidate character group. Among the above, the highest connection probability is a connection probability between the one recognition candidate character and the immediately preceding or immediately following recognition candidate character group.
[0011]
According to a second aspect of the present invention, in the first aspect, the feature extraction unit is configured such that each recognition candidate character in the determination target recognition candidate character group and a recognition candidate character having the highest certainty factor in the immediately preceding recognition candidate character group. A connection probability value between the recognition candidate characters of the determination target recognition candidate character group, and a value of a connection probability between the recognition candidate character of the highest certainty factor in the immediately following recognition candidate character group And
[0013]
According to a third aspect of the present invention, in the first or second aspect, the feature extraction unit is configured to determine the determination target based on the value of the connection probability and the certainty value of each recognition candidate character in the determination target recognition candidate character group. A feature amount of the recognition candidate character group is extracted.
[0014]
According to a fourth aspect of the present invention, in any one of the first to third aspects, the feature extraction unit is configured to determine whether the determination target recognition candidate character group and the immediately subsequent recognition candidate character group when the immediately preceding recognition candidate character group does not exist. If the immediately following candidate character group does not exist, the connection probability between the immediately preceding candidate character group and the immediately preceding candidate candidate character group is substituted. The connection probability is substituted as the connection probability between the immediately-recognized candidate character group.
[0015]
According to a fifth aspect of the present invention, in any one of the first to fourth aspects, the reliability calculation means calculates a probability of one recognition candidate character in the determination target recognition candidate character group from the feature amount as a determination score. A determination score calculating means for calculating the reliability based on the determination score.
[0016]
According to a sixth aspect of the present invention, in any one of the first to fifth aspects, the post-processing control means limits recognition candidate characters to be post-processed based on the reliability calculated from the reliability calculation means. It is characterized by doing.
[0017]
The invention of claim 7 is a character recognition reliability determination method for determining the reliability of one candidate character based on a series of candidate character concatenation relations, and each candidate character in the determination target candidate character group and the immediately preceding character character A feature extraction step of extracting a value of the connection probability between the candidate character group and / or a value of the connection probability between the candidate character group immediately after it as a feature amount of the determination target candidate character group; and the feature extraction step right reading of the sample data first of the recognition candidate characters in the concatenated feature vector of feature value the determination target candidate character group is a positive read a concatenated feature vector is extracted in advance is known to be a misreading from And a reliability calculation step of calculating the reliability of the candidate character group for determination based on the distribution of each set of misreads .
Furthermore, the feature extraction step includes a connection probability between one candidate character in the determination target candidate character group and each candidate character from the first place to the M place in the candidate character group immediately before or immediately after the candidate character group, The highest connection probability is a connection probability between the one candidate character and the immediately preceding or immediately following candidate character group.
[0018]
The invention according to claim 8 is the method according to claim 7 , wherein the feature extracting step is a connection between each candidate character in the determination target candidate character group and a candidate character of the highest certainty factor in the immediately preceding candidate character group. The value of the probability and / or the value of the connection probability with the candidate character of the highest certainty factor in the immediately following candidate character group is extracted as the feature amount of the determination target candidate character group.
[0020]
The invention according to claim 9 is the determination target candidate character according to claim 7 or 8 , wherein the feature extraction step is based on the value of the connection probability and the confidence value of each candidate character in the determination target candidate character group. A feature amount of the group is extracted.
[0021]
According to a tenth aspect of the present invention, in any one of the seventh to ninth aspects, the feature extraction step includes the determination target candidate character group and the immediately following candidate character group when the immediately preceding candidate character group does not exist. If the immediately following candidate character group does not exist, the connection probability between the previous candidate character group and the immediately preceding candidate character group is substituted. The connection probability between them is substituted as the connection probability between the immediately following candidate character group.
[0022]
According to an eleventh aspect of the present invention, in any one of the seventh to tenth aspects, the reliability calculation step calculates a probability of one candidate character in the determination target candidate character group from the feature amount as a determination score. A score calculation step is included, and the reliability is calculated based on the discrimination score.
[0023]
The invention of claim 12 is a program for executing each processing step in the character recognition reliability determination method according to any one of claims 7 to 11 .
[0024]
A thirteenth aspect of the present invention is a recording medium storing a program for executing each processing step in the character recognition reliability determination method according to any one of the seventh to eleventh aspects .
[0025]
The features of the present invention will become more apparent from the following description of embodiments.
[0026]
The “character recognition means” in the claims corresponds to the character recognition unit 2 and the character recognition dictionary 3 in the embodiment. The “feature extraction means” in the claims corresponds to the connection feature extraction unit 4 and the character connection probability dictionary 5 in the embodiment. The “reliability calculation means” in the claims corresponds to the discrimination score calculation unit 6, the recognition reliability calculation unit 7, and the discrimination score-reliability conversion table 8 in the embodiment. The “post-processing control means” in the claims corresponds to the recognition candidate number control unit 9 and the reliability-cumulative correct reading rate table 10 in the embodiment.
[0027]
A “feature extraction step” in the claims corresponds to step S103 in FIG. 7 in the embodiment. The “reliability calculation step” in the claims corresponds to steps S104 and S105 in FIG. 7 in the embodiment.
[0028]
However, the following embodiment is merely one embodiment of the present invention, and the meaning of the term of the present invention or each constituent element is not limited to that described in the following embodiment. Absent.
[0029]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below with reference to the drawings.
[0030]
First, FIG. 1 shows a circuit block diagram of a handwritten character recognition apparatus according to an embodiment. In the figure, reference numeral 1 denotes an input unit which generates and outputs handwritten character information from handwriting input by handwriting on a tablet or the like. Reference numeral 2 denotes a character recognition unit which compares handwritten character information supplied from the input unit 1 with the character feature amount of the character recognition dictionary 3, and the degree of proximity (confidence) between the two in the recognition dictionary from the first to the Nth. The character is output as a recognition candidate character for the handwritten character. Reference numeral 3 denotes a character recognition dictionary in which candidate characters are stored together with their character feature amounts.
[0031]
Reference numeral 4 denotes a connected feature extraction unit, which extracts the connected feature amount of one recognition candidate character that is the object of correctness determination from the recognition candidate character string supplied from the character recognition unit 2 with reference to the intercharacter connection probability dictionary 5. To do. Reference numeral 5 denotes an inter-character concatenation dictionary, which stores the probability that two predetermined characters are concatenated in association with the two characters.
[0032]
A discrimination score calculation unit 6 processes the connected feature amount from the connection feature extraction unit 4, and calculates a correct / incorrect discrimination score of the recognition candidate character group to be determined.
[0033]
A recognition reliability calculation unit 7 compares the discrimination score from the discrimination score calculation unit 6 with the discrimination score-reliability conversion table 8 and outputs the reliability of the recognition candidate character group to be determined. A discrimination score-reliability conversion table 8 stores the relationship between the discrimination score and the reliability as a table.
[0034]
A recognition candidate number control unit 9 compares the reliability from the recognition reliability 7 with the reliability-cumulative correct reading rate table 10 and limits the number of recognition candidates for the determination target character. Reference numeral 10 denotes a reliability-cumulative correct reading rate table which stores the relationship between reliability and correct reading rate as a table.
[0035]
A post-processing unit 11 performs post-processing on the number of recognition candidate characters set by the recognition candidate number control unit 9. For example, in the example of FIG. 9, the number of recognition candidates for the written character “ko” is M1, and the written characters “to”, “bo”, “not”, “possible”, “missing”, “de” Assuming that the number of recognition candidates for M2, M3, M4, M5, M6, and M7 is set by the recognition candidate number control means 9, each of the recognition candidate characters generated by the character recognition unit 2 is certain. The post-processing unit 11 performs post-processing on only the recognition candidate characters having the upper ranks M1, M2, M3, M4, M5, M6, and M7.
[0036]
Next, details of processing of each unit shown in the circuit block diagram will be described. First, the process of the connected feature extraction unit 4 will be described with reference to FIG. In the connected feature extraction unit 4, any one of the connection probability Pbk between the recognition candidate character group to be determined and the first candidate character immediately before it, and the connection probability Pfk between the first candidate character immediately after the connection candidate character group. Either one or both are read from the inter-character concatenation probability dictionary 5 and used as the connection feature amount of the recognition candidate character group.
[0037]
For example, in the case of extracting the connection feature amount using both the connection probabilities Pbk and Pfk, as shown in FIG. 2, the recognition candidate character group (from the first confidence level to the Nth rank) of the determination target character is “ Connection probability Pb1, Pb2,..., PbN between the eyes “day”,. The connection probabilities Pf1, Pf2,..., PfN between the first character “ha” are read from the inter-character connection probability dictionary 5, and the values of these connection probabilities are used as the connection feature amount of the recognition candidate character group to be determined. To do.
[0038]
In the above description, the connected feature amount of the recognition candidate character group is extracted by both Pbk and Pfk, but may be extracted by either Pbk or Pfk. In addition, in the case of extracting by both Pbk and Pfk, when there is no immediately preceding first recognition candidate character (for example, when the determination target character is the head of the character string), the candidate character A connection feature amount is extracted by substituting the connection probability Pfk for the immediately following character as it is with the connection probability Pbk for the immediately preceding character. Similarly, when there is no immediately following first recognition candidate character (for example, when the determination target character is the end of the character string), Pbk is used as it is.
[0039]
Next, with reference to FIG. 3, the process in the discrimination score calculation unit 6 will be described. The values of Pbk and Pfk extracted by the connected feature extraction unit 4 can be expressed as a predetermined vector (connected feature vector) in a 2N-dimensional vector space. The discrimination score calculation unit 6 extracts a connected feature vector as learning data for a sample in which the first recognition candidate character is correctly read or misread in advance, and the connected feature vector of the recognition candidate character group to be determined and Are compared, and the discrimination score of the recognition candidate character group is calculated.
[0040]
For example, it is assumed that the learning data of correctly connected and misread connected feature vectors and the connected feature vectors of the recognition candidate character group to be determined are in a state as shown in FIG. The discrimination score calculation unit 6 obtains an average value (centroid vector) and a covariance matrix of feature vectors of both classes in advance from the distribution of feature vectors of the correct reading and misreading sets (classes), and stores them. ing. Then, the Mahalanobis distances DMc and DMe between the center of gravity of each class and the connected feature vector of the character group to be determined are obtained, and the ratio of these values, the logarithm of the ratio or the difference is used as a discrimination score.
[0041]
Here, the Mahalanobis distance DM is calculated as follows. That, _{m 1} centroid vectors Class _{C 1,} when the covariance matrix of class _{C 1} and sigma _1, Mahalanobis squared distance DM ₁ to _{m 1} from the predetermined feature vector x is defined by the following equation.
[0042]
DM ₁ = (x−m ₁ ) ^t Σ1 ⁻¹ (x−m ₁ ) (t: transpose, −1: inverse matrix) where the covariance matrix Σ ₁ is an n × n square matrix (n: The number of dimensions of the feature space), and its (i, j) element is the covariance of the i-th feature quantity and the j-th feature quantity, that is, Σ ₁ (i, j) = σ _ij .
[0043]
In the above, the discriminant score is calculated using the Mahalanobis distance DM. Instead, a linear discriminant function is obtained by linear discriminant analysis from the distribution of feature vectors of the correct reading and misreading classes, and the determination target is obtained. A discrimination score may be obtained by applying this linear discriminant function to the feature vector of the recognition candidate character group. In addition, using the neural network trained so that the feature vector extracted from the correct reading and misreading learning samples can be determined whether the target feature vector is correct reading or misreading as the learning data, the neural network for the determination target feature vector is used. The net output value may be used as the discrimination score.
[0044]
Next, processing of the recognition reliability calculation unit 7 will be described with reference to FIG. For example, in the calculation of the discrimination score, when the Mahalanobis distance DMc, DMe distance ratio or the logarithm of the ratio is used as the discrimination score, the relationship between the discrimination score obtained from each of the correctly read and misread learning samples and the reliability is shown in FIG. As shown. Here, the reliability is calculated from the learning sample by Bayes' theorem.
[0045]
That is, the ratio of the number of correctly read samples having the discrimination score y to the total number of correctly read samples is p (y | X1 = C), and the ratio of the number of misread samples having the discrimination score y to the number of all misread samples is p (y | X1 = E), the ratio of the total number of correctly read samples to the total number of samples is P (X1 = C), and the ratio of the total number of misread samples to the total number of samples is P (X1 = E), it has a discrimination score y The reliability of the recognition candidate character having the highest certainty in the recognition candidate character group can be calculated by the following equation.
[0046]
P (X1 = C | y) = p (y | X1 = C) · P (X1 = C) / [p (y | X1 = C) · P (X1 = C) + p (y | X1 = E) · P (X1 = E)]
Here, X1 represents the first recognition candidate character, and X1 = C and X1 = E mean events where the first recognition candidate character is correct and incorrect, respectively.
[0047]
A discriminant score-reliability conversion table showing the relationship between the discriminant score and the reliability is created in advance from this equation, and this is stored in the discriminant score-reliability conversion table 8. The recognition reliability calculation unit 7 compares the discrimination score from the discrimination score calculation unit 6 with the score of the discrimination score-reliability conversion table 8, and determines the corresponding reliability as the first recognition of the recognition candidate character group. Output as reliability of candidate characters.
[0048]
Next, the processing of the recognition candidate number control unit 9 will be described with reference to FIG. The table shown in the upper part of FIG. 5 shows the reliability of the recognition candidate character with the first confidence level of the determination target and the cumulative probability that the correct recognition character is included in the recognition candidate characters up to the Nth determination target. It shows the relationship. The probabilities in the table are calculated in advance based on the correct reading and misreading learning samples.
[0049]
The reliability-cumulative correct reading rate table 10 stores such a table. Then, the recognition candidate number control unit 9 compares the reliability of the recognition candidate character group to be determined with the reliability level in the table, and recognizes up to what level while referring to the cumulative probability of the corresponding reliability level. Whether to output the candidate character to the language processing unit 11 is determined. Here, how much is output is determined by, for example, whether or not the cumulative probability of the corresponding reliability level has reached a predetermined threshold value. At this time, the set threshold value may be uniform for all reliability levels, or may be set individually for each reliability level.
[0050]
Alternatively, the number of output candidates for each reliability level may be set in advance based on the upper table in FIG. 5 and stored in the reliability-cumulative correct reading rate table 10. The table shown in the lower part of FIG. 5 is an example when the reliability level and the number of output candidates are preset. When the table is stored in the reliability-cumulative correct reading rate table 10 in advance, the recognition candidate number control unit 9 reads the corresponding output candidate number from the table and outputs it to the language processing unit 11 accordingly. Restrict recognition candidate characters.
[0051]
In the above embodiment, instead of individually comparing the connection probabilities for each recognition candidate character, the reliability of the recognition result is calculated based on the statistical tendency of the combination probability value of the recognition candidate character group. Because this sets the number of recognition candidate outputs, even if there are individual connection relationships that are correctly read but have a low connection probability, and those that are misread but have a high connection probability, such a connection relationship is used for reliability determination. It is possible to suppress the influence, so that recognition candidate characters having a high correct reading rate can be output to the language processing unit.
[0052]
Next, a second embodiment according to the present invention will be described below. In the present embodiment, the connection feature amount extraction processing in the connection feature extraction unit 2 is changed. That is, in the above embodiment, as the connection probability (Pbk, Pfk) of the recognition candidate character to be determined, the connection probability between the recognition candidate character to be determined and the first character immediately before or the first character immediately after is determined. In this embodiment, the maximum connection probability between the recognition candidate character to be determined and the immediately preceding first to Mth characters is Pbk. Similarly, the recognition candidate character to be determined Let Pfk be the maximum value of the connection probability between the immediately preceding first to Mth characters.
[0053]
For example, in the example of FIG. 6, Pb1 for the first character “day” of the recognition candidate character to be determined is the concatenation of the “day” and the immediately preceding first character “直前” to the M character “hu”. Among the probabilities P (C1 | Cbk), the maximum connection probability is adopted. Further, Pf1 for the first character “day” is the largest concatenation probability among the respective connection probabilities P (Cfk | C1) from the first character “mo” immediately after that “day” to the M character “dead”. Adopt probability. Similarly, Pb2 and Pf2 for the second character “Month” of the recognition candidate character to be determined adopt the maximum values of the connection probabilities for the immediately preceding and immediately following character groups, respectively.
[0054]
Here, C1 represents the character at the first recognition candidate to be determined, and Cbk and Cfk represent the character at the kth recognition candidate immediately before and after, respectively. P (Cj | Ci) represents the connection probability that the character Cj appears after the character Ci.
[0055]
In the second embodiment, since the maximum connection probability between the immediately preceding and immediately following character groups is adopted as Pbk and Pfk, in addition to the function and effect in the first embodiment, Even when the first recognition candidate character immediately after is incorrect, highly accurate reliability determination can be performed.
[0056]
As yet another embodiment, in addition to the above connection probabilities Pbk and Pfk, the certainty factor (similarity or distance value) of recognition candidate characters up to the Lth position is added as a feature element, and the relevant factor is expressed in a (2N + L) -dimensional vector space. You may make it extract the feature vector of a recognition candidate character group. In such a case, the discrimination space shown in FIG. 3 has (2N + L) dimensions. In addition, the correct reading / misreading samples are also determined based on the sample data by using the connection probability Pbk, Pfk for the recognition candidate character group of the sample and the certainty of the recognition candidate characters up to the L position as feature quantity extraction elements. Tables stored in the score-reliability conversion table 8 and the reliability-cumulative correct reading rate table 10 are set. In the third embodiment, since not only the connection relationship but also the certainty factor is taken into account, it is possible to determine the reliability with higher accuracy.
[0057]
Incidentally, in the above embodiment, a series of processing flow has been described by dividing the processing for each block in FIG. 1, but it is also possible to execute such processing flow by the CPU according to the control program. In such a case, the processing flow is stored as a control program in the ROM or RAM. Further, reference data of the character recognition dictionary 3, the inter-character connection probability dictionary 5, the discrimination score-reliability conversion table 8, and the reliability-cumulative correct reading rate table 10 are also stored in the ROM or RAM. The CPU executes the above processing according to the control program while referring to the reference data.
[0058]
FIG. 7 shows a flow according to such a control program. Here, step S101 is processing in the input unit 1, step S102 is processing in the character recognition unit 2, step S103 is processing in the connected feature extraction unit 4, step S104 is processing in the discrimination score calculation unit 6, and step S105. Is a process in the recognition reliability calculation unit 7, and step S106 is a process in the recognition candidate number control unit 9.
[0059]
Such control program and various reference data can be traded via a recording medium such as a floppy disk or a transmission medium such as the Internet. An example of a file structure of data traded through a recording medium or a transmission medium is shown in FIG. Data having such a file structure is recorded on the recording medium. In a transaction via a transmission medium, data having such a file structure is supplied via the transmission medium.
[0060]
However, if only the reliability of the candidate character to be determined is determined, it is not necessary to supply data relating to all the steps shown in FIG. 7, and among these, an execution program for steps S103 to S105 and a program required for it are required. It is sufficient to supply only the inter-character concatenation probability dictionary data and the discrimination score-reliability conversion table data. In such a case, these data are stored as file data in the supplied recording medium. When the data is supplied via a transmission medium, the file data is transmitted.
[0061]
The embodiment according to the present invention has been described above. However, the present invention is not limited to the embodiment, and various other modifications are possible. For example, in the above-described embodiment, an example in which the reliability determination method according to the present invention is implemented in a handwritten character recognition device has been described. However, the reliability determination method according to the present invention is not limited to this, and is optical. The present invention can be widely applied to a character recognition device or a speech recognition device.
[0062]
In the above embodiment, the feature of the reliability determination method according to the present invention lies in steps S103 to S105 in FIG. 7, and the connection relationship between the immediately preceding character and the immediately following character or either one of the characters is determined. The determination score is calculated according to the distribution state, and the reliability of the target character is determined based on the determination score.
[0063]
Therefore, the processes in steps S103 to S105 can be used not only in the handwritten character recognition apparatus according to the above embodiment but also in other recognition apparatuses. For example, steps S101 and S102 in FIG. 7 may be a known optical character recognition device or a recognition process by voice input. In step S106, the recognition result may be rejected (invalidated) according to the degree of reliability as well as the number of recognition candidates.
[0064]
In addition, the determination score calculation method in step S104 and the recognition reliability calculation method in step S105 also employ methods other than the method using the Mahalanobis distance DM and the method using Bayes' theorem shown in the above embodiment. You can also
[0065]
The embodiment of the present invention can be variously modified as appropriate within the scope of the technical idea of the present invention.
[0066]
【The invention's effect】
As described above, according to the present invention, even when the concatenation probability for correct reading and misreading is inappropriate, or when the candidate character with the highest degree of certainty immediately before and immediately after is incorrect, the reliability of the determination target character is relatively accurate. Therefore, when this is adopted in the character recognition device, the accuracy of the recognition candidate character can be improved.
[Brief description of the drawings]
FIG. 1 is a diagram showing a circuit block diagram according to an embodiment. FIG. 2 is a diagram for explaining processing of a connected feature extracting unit according to the embodiment. FIG. 3 is a diagram of a discrimination score calculating unit according to the embodiment. FIG. 4 is a diagram for explaining the processing. FIG. 4 is a diagram for explaining the processing of the recognition reliability calculation unit according to the embodiment. FIG. 5 is a diagram for explaining the processing of the recognition candidate number control unit according to the embodiment. FIG. 6 is a diagram for explaining another process of the connected feature extraction unit according to the embodiment. FIG. 7 is an execution flowchart according to the embodiment. FIG. 8 is an execution program and reference data according to the embodiment. File structure [Fig. 9] Diagram for explaining a conventional example [Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Input part 2 Character recognition part 3 Character recognition dictionary 4 Concatenated feature extraction part 5 Inter-character connection probability dictionary 6 Discrimination score calculation part 7 Recognition reliability calculation part 8 Discrimination score-reliability conversion table 9 Recognition candidate number control part 10 Reliability -Accumulated correct reading rate table

Claims

Character recognition means for processing input information and outputting a recognition candidate character group;
The value of the connection probability between each recognition candidate character in the determination target recognition candidate character group for one input information output by the character recognition means and the immediately preceding recognition candidate character group for the input information immediately before and / or immediately after that Feature extraction means for extracting the value of the connection probability between the immediately-recognized candidate character group for the input information as the feature amount of the determination-target recognition candidate character group;
A concatenated feature vector composed of feature amounts from the feature extraction means and a pre-extracted concatenated feature vector in which it is known that the first recognition candidate character of the determination target recognition candidate character group is correct or misread. Based on the distribution of each set of correct reading and misreading of certain sample data, reliability calculation means for calculating the reliability of the determination target recognition candidate character group, and based on the reliability from the reliability calculation means have a recognition candidate number control unit to determine the number of recognition candidate characters is output from the determination target recognition candidate character group as a post-treatment of the subject,
The feature extraction means includes a probability of concatenation between one recognition candidate character in the determination target recognition candidate character group and each recognition candidate character from the first position to the M position in the immediately preceding or immediately following recognition candidate character group. A character recognition apparatus characterized in that the highest connection probability is a connection probability between the one recognition candidate character and the immediately preceding or immediately following recognition candidate character group.

The feature extraction unit according to claim 1, wherein the feature extraction unit includes a connection probability value between each recognition candidate character in the determination target recognition candidate character group and a recognition candidate character having the highest certainty factor in the immediately preceding recognition candidate character group, and A character recognition apparatus that extracts a value of a connection probability with a recognition candidate character with the highest certainty factor in the immediately following recognition candidate character group as a feature amount of the determination target recognition candidate character group.

3. The feature amount of the determination target recognition candidate character group according to claim 1 or 2 , wherein the feature extraction means uses the value of the connection probability and the confidence value of each recognition candidate character in the determination target recognition candidate character group. A character recognition device characterized by extracting a character.

4. The feature extraction unit according to claim 1 , wherein when the immediately preceding recognition candidate character group does not exist, the feature extraction unit calculates a connection probability between the determination target recognition candidate character group and the immediately following recognition candidate character group. Substituting as the connection probability between the recognition candidate character group and, if the immediately following recognition candidate character group does not exist, the connection probability between the determination target recognition candidate character group and the immediately preceding recognition candidate character group is the immediately following recognition candidate A character recognition device that substitutes as a connection probability between character groups.

5. The discrimination score calculation unit according to claim 1 , wherein the reliability calculation unit includes a discrimination score calculation unit that calculates a probability of one recognition candidate character in the determination target recognition candidate character group as a discrimination score from the feature amount. A character recognition device that calculates the reliability based on the discrimination score.

6. The recognition candidate number control unit according to claim 1 , wherein the recognition candidate number control unit restricts recognition candidate characters to be post-processed based on the reliability calculated by the reliability calculation means. Character recognition device.

A character recognition reliability determination method for determining the reliability of one candidate character by a series of candidate character connection relations,
A value of the connection probability between each candidate character in the determination target candidate character group and the immediately preceding candidate character group and / or a connection probability value between the candidate character group immediately after the candidate character group and the determination target candidate character group A feature extraction step of extracting as a feature quantity of
A concatenated feature vector composed of feature amounts from the feature extraction step , and a pre-extracted concatenated feature vector in which it is known that the first recognition candidate character of the determination target candidate character group is correct or misread. A reliability calculation step of calculating the reliability of the determination target candidate character group based on the distribution of each set of correct reading and misreading of the sample data;
I have a,
In the feature extraction step, the highest probability of connection between one candidate character in the candidate character group to be determined and each candidate character from the first to the M-th candidate character immediately before or immediately after the candidate character group. A character recognition reliability determination method, wherein a connection probability is a connection probability between the one candidate character and the immediately preceding or immediately following candidate character group.

8. The feature extraction step according to claim 7 , wherein the feature extraction step includes a connection probability value between each candidate character in the determination target candidate character group and a candidate character of the highest certainty factor in the immediately preceding candidate character group and / or the A character recognition reliability determination method characterized by extracting a value of a connection probability with a candidate character of the highest certainty level immediately after a candidate character group as a feature amount of the determination target candidate character group.

9. The feature extraction step according to claim 7 , wherein the feature extraction step extracts a feature amount of the determination target candidate character group from the value of the connection probability and the confidence value of each candidate character in the determination target candidate character group. Character recognition reliability determination method characterized by the above.

In any one of Claims 7 thru | or 9 , The said feature extraction step WHEREIN: When the said previous candidate character group does not exist, the connection probability between the said determination object candidate character group and the said next candidate character group is said If the immediately preceding candidate character group does not exist, the connection probability between the determination target candidate character group and the immediately preceding candidate character group is used as the immediately preceding candidate character group A character recognition reliability determination method characterized by substituting as a probability of connection with a candidate character group.

In any one of Claims 7 thru | or 10 , The said reliability calculation step includes the discrimination | determination score calculation step which calculates the probability of one candidate character in the said determination target candidate character group as a discrimination | determination score from the said feature-value, A character recognition reliability determination method, wherein the reliability is calculated based on a discrimination score.

The program which performs each process step in the character recognition reliability determination method in any one of Claims 7 thru | or 11 .

The recording medium which memorize | stored the program which performs each process step in the character recognition reliability determination method in any one of Claims 7 thru | or 11 .