JP3560794B2

JP3560794B2 - Character image normalization method and character recognition device

Info

Publication number: JP3560794B2
Application number: JP34231397A
Authority: JP
Inventors: 寿太郎石岡
Original assignee: 株式会社日本デジタル研究所
Priority date: 1997-10-31
Filing date: 1997-10-31
Publication date: 2004-09-02
Anticipated expiration: 2017-10-31
Also published as: JPH11134435A

Description

【０００１】
【発明の属する技術分野】
本発明は、文字認識装置に関し、特に、記入線の太さや文字の大きさが異なる文字画像を正規化する技術に関する。
【０００２】
【従来の技術】
文字認識装置では、読取った文字画像の線の太さは一様とはいえず、文字の大きさも一定ではなく、様々な線の太さと大きさの文字画像が得られる。このような文字画像から文字認識を行なうために、前処理として読取った文字画像の正規化を行なっている。
【０００３】
読取られた文字画像の線の太さと大きさを正規化する方法として、間引き正規化技法およびＯＲ正規化技法が知られている。ここで、間引き正規化は、画素を一定の間隔で抽出し、他の画素を切捨てることにより入力画素より小さいサイズの画素を生成する方法であり、ＯＲ正規化は一定間隔内の画素のＯＲ（論理和）をとることにより、入力画像より小さいサイズの画像を生成する方法である。
【０００４】
しかしながら、間引き正規化の場合には細かい線が消失したり、かすれたりする等の減少が生じやすいという問題点があり、また、ＯＲ正規化では文字のループ内の画素が潰れやすい等の問題点があった。
【０００５】
上記問題点を考慮したものとして、入力画像に対して線分の細め処理を行ない、細め処理後の画像に対してＯＲ正規化を行なう方法（特開平３−３７７８３号）がある。
【０００６】
更に、上記特開平３−３７７８３号に開示の技術を改善したものとして、入力画像に対して線分の細め処理をその強さを変えて施し、それぞれの細め処理後の文字画像に対してＯＲ正規化することにより、一つの画像に対して複数の正規化文字画像を得てそれぞれ認識処理を行ない、得られた複数の認識結果から最終認識結果を得る方法（特開平４−６８７号）がある。
【０００７】
【発明が解決しようとする課題】
上記特開平４−６８７号に開示の文字認識技法では、入力画像に施す線分の細め処理の強さを変えて、それぞれの細め処理後の文字画像に対してＯＲ正規化を行なうので、一文字の中で線分の太さの変化の程度が大きい場合にはその度合いに応じて細め処理の回数が増大することになり、ＯＲ正規化による正規化文字も増大し、認識処理の回数が増す結果になり、線分の幅の変化が一定以上の場合には処理速度が遅くなる等、メモリー容量および認識処理速度の上から一定の制限があった。
【０００８】
例えば、ボールペンで書かれた文字を例にとるとボールペンではペン先にインクがたまることがままあり、そのままの状態で記入を続けていると、一文字の中でも記入線が太くなったり細くなったりして文字中の線分の太さの変化が大きくなる場合がある。また、複写等で拡大した文字の場合、倍率が大きい場合と線幅が太くなり、画質が粗くなる。
【０００９】
筆記具にも鉛筆、ボールペン、サインペン等、様々なものがあり芯の太さにも各種の種類がある。このように筆記具によって書かれる文字のあらゆる線幅の文字を同一のパターンマッチングによる認識処理で行なう際に、線幅によって特徴データが変る特徴抽出法を用いると線幅に応じた多くの辞書を要することとなり、認識処理の効率が低下するという問題点があった。
【００１０】
本発明は、上記従来技術の問題点に鑑みてなされたものであり、一文字中の線の太さの変動があっても線幅に応じた辞書を要することなく、安定した認識を可能とする文字画像の正規化方法および正規化処理手段を備えた文字認識装置の提供を目的とする。
【００１１】
【課題を解決するための手段】
上記課題を解決するために、入力画像から文字画像を切出して正規化処理を行なった後、正規化された文字画像に対して文字認識処理を行なう文字認識装置において、切出された文字画像の外接矩形を検出する工程と、外接矩形の大きさと外接矩形に占める文字の画素数の割合から、切出された文字画像に対して、拡大正規化のみ、間引き正規化のみ、または細線化のいずれを行なうかを判定する工程と、上記判定結果に基づいて、文字画像の拡大正規化処理のみ、間引き正規化処理のみ、または細線化処理のいずれかを行なう工程と、を有し、さらに、上記細線化処理の工程の後段に、細線化処理後の文字画像の外接矩形を検出してその大きさにより当該文字画像を拡大正規化するか縮小正規化するかを判定する工程と、上記判定結果に基づいて細線化された文字画像の拡大正規化処理または縮小正規化処理を行なう工程と、上記、拡大正規化処理または縮小正規化処理された文字画像に対して太め処理を施して所定の太さの線幅の文字画像とする工程、とを有することを特徴とする。
【００１２】
また、第２の発明の文字画像の正規化方法は、入力画像から文字画像を切出して正規化処理を行なった後、正規化された文字画像に対して文字認識処理を行なう文字認識装置において、切出された文字画像の外接矩形を検出する工程と、外接矩形の大きさと外接矩形に占める文字の画素数の割合から、切出された文字画像に対して、拡大正規化のみを行なうか、間引き正規化のみを行なうか、拡大正規化、間引き正規化、または縮小化のいずれかを行った後に細線化を行なうかを判定する工程と、上記判定結果に基づいて、文字画像の拡大正規化処理のみ、または間引き正規化処理のみを行なう工程と、上記判定結果に基づいて、文字画像の拡大正規化処理、間引き正規化処理、または縮小化処理のいずれかを行った後に細線化処理を行う工程と、この細線化処理により細線化された文字画像に太め処理を施して所定の太さの線幅の文字画像とする工程、とを有することを特徴とする。
【００１３】
また、第３の発明の文字認識装置は、入力画像から文字画像を切出して正規化処理を行なった後、正規化された文字画像に対して文字認識処理を行なう文字認識装置において、切出された文字画像の外接矩形を検出する第１の外接矩形検出手段と、外接矩形の大きさと外接矩形に占める文字の画素数の割合から、切出された文字画像に対して、拡大正規化のみ、間引き正規化のみ、または細線化のいずれを行なうかを判定する文字画素検出手段と、上記文字画素検出手段の判定結果に基く文字画像の拡大正規化処理、又は第２の外接矩形検出手段の判定結果に基づく細線化処理後の文字画像の拡大正規化処理を行なう拡大正規化手段と、上記文字画素検出手段の判定結果に基づいて文字画像の間引き正規化処理を行なう間引き正規化手段と、上記文字画素検出手段の判定結果に基づいて文字画像の細線化処理を行なう細線化処理手段とを有し、さらに、細線化処理後の文字画像の外接外接矩形を検出してその大きさにより当該文字画像を拡大正規化するか縮小正規化するかを判定する第２の外接矩形検出手段と、上記第２の外接矩形検出手段の判定結果に基づいて細線化処理後の文字画像の縮小正規化処理を行う縮小正規化手段と、上記第２の外接矩形検出手段の判定結果に基づいて、拡大正規化手段により拡大正規化された文字画像又は縮小正規化手段により縮小正規化された文字画像に対して太め処理を施して所定の太さの線幅の文字画像とする太め処理手段と、を有する、文字正規化手段を備えたことを特徴とする。
【００１４】
また、第４の発明の文字認識装置は、入力画像から文字画像を切出して正規化処理を行なった後、正規化された文字画像に対して文字認識処理を行なう文字認識装置において、切出された文字画像の外接矩形を検出する外接矩形検出手段と、外接矩形の大きさと外接矩形に占める文字の画素数の割合から、切出された文字画像に対して、拡大正規化のみを行うか、間引き正規化のみを行うか、拡大正規化、間引き正規化、または縮小化のいずれかを行った後に細線化を行うかを判定する文字画素検出手段と、上記文字画素検出手段の判定結果に基づいて文字画像の拡大正規化処理を行なう拡大正規化手段と、上記文字画素検出手段の判定結果に基づいて文字画像の間引き正規化処理を行なう間引き正規化手段と、上記文字画素検出手段の判定結果に基づいて文字画像の縮小正規化処理を行なう縮小正規化手段とを有し、さらに、上記文字画素検出手段の判定結果に基づいて、拡大正規化手段による拡大正規化処理、間引き正規化処理手段による間引き正規化処理、または縮小正規化手段による縮小正規化処理のいずれかを行った後の文字画像に対して細線化処理を行なう細線化処理手段と、上記細線化処理手段により細線化処理された文字画像に太め処理を施して所定の太さの線幅の文字画像とする太め処理手段と、を有する、文字正規化手段を備えたことを特徴とする。
【００１５】
【００１６】
【００１７】
【００１８】
【００１９】
【発明の実施の形態】
本発明の文字画像の正規化方法は、切出された文字画像の線分の外接矩形を検出し、外接矩形の大きさにより拡大、縮小、間引、または細線化による正規化処理等で線分の幅をほぼ一定の太さにする。
【００２０】
＜実施形態１＞
本発明の正規化方法の第１は、切出された文字画像の線分の細線化処理を行なった後、外接矩形を検出し、外接矩形の大きさにより拡大正規化または縮小正規化を行ない、得られた線分に太め処理を施して線分の幅を一定にする。
【００２１】
図１は、上記第１の正規化方法を適用した文字認識装置の一実施例の構成図であり、文字認識装置１０は文字切り出し部１１０、正規化処理ブロック１２０、および文字認識部１３０から構成され、正規化処理ブロック１２０は細線化処理部１２１、外接矩形検出部１２２、拡大正規化部１２３、ＯＲ正規化部１２４、および太め処理部１２５を有している。
【００２２】
文字切り出し部１１０はスキャナー等の文字読取り装置で読み込んだ画像Ｉｍから一文字ずつ文字画像を切出す。
【００２３】
正規化処理ブロック１２０は、切出された一文字分の画像の線分の細線化処理を行なった後、外接矩形を検出し、外接矩形の大きさにより拡大正規化または縮小正規化を行なってサイズを一定にし、得られた線分の幅を予め設定された太さにして認識処理部１３０に与える。そのために、先ず、細線化処理部１２１で文字切り出し部１１０によって切出された１文字分の文字画像の細線化処理を行ない、その細線化画像を得る。次に、細線化処理部１２１で細線化された画像に対して外接矩形検出部１２２でその文字の外接矩形を検出し、外接矩形の大きさによって、当該文字画像の拡大または縮小判定を行なう。すなわち、外接矩形検出部１２２では外接矩形の大きさをｎ×ｍドットとし、ｎ'×ｍ'を文字画像の正規化サイズとする時、ｎおよびｍがｎ'より小さく、かつｎおよびｍがｍ'より小さいとき、その文字画像に拡大処理が必要と判定し、ｎまたはｍの少なくとも一方がｎ'或いはｎより大きい時、その文字画像に縮小処理が必要と判定する。
【００２４】
外接矩形検出部１２２で拡大判定がなされた場合には、拡大正規化部１２３で細線化処理部１２２によって細線化された文字画像を拡大正規化処理してサイズが正規化された画像を得る。また、縮小判定がなされた場合には、ＯＲ正規化部１２４で細線化処理部１２２によって細線化された文字画像をＯＲ正規化処理してサイズが正規化された画像を得る。
【００２５】
拡大正規化部１２３またはＯＲ正規化部１２４でサイズが正規化された文字画像に対して太め処理部１２５で文字画像の線分の幅が予め設定された太さになるように太め処理を施す。
【００２６】
文字認識部１３０は、正規化処理ブロック１２０で生成された正規化画像をもとに文字の認識を行ない、その文字コードを認識結果として出力する。
【００２７】
［実施例］
本実施形態の正規化ブロック１２０の一実施例について以下説明する。なお、図５は入力画像の例を示す図、図６は図５の画像から切出された文字画像の例を示す図、図７は細線化処理された文字画像の例を示す図、図８は拡大正規化された文字画像の例を示す図、図９は正規化された文字画像の例を示す図である。スキャナー等から読み込んだ画像Ｉｍ（図５）を１文字分ずつ切り出して線幅が３ドット、大きさが３０×３０ドットの文字画像（図９）に正規化する場合を例とし、文字切り出し部１２０によって切出された５文字分の文字画像のうち、ボールペンのペン先にインクが溜まったため書き始めが太くて途中がかすれている'２'の文字画像（図６）を正規化する場合について述べる。
【００２８】
正規化ブロック１２０に取り込まれた文字'２'を現わす文字画像６１（図６）は、先ず、細線化処理部１２１で細線化処理され、図７に示すような細線化文字画像となる。次に、外接矩形検出部１２２で細線化された文字画像７１の外接矩形（図７）を求める（図７の例では、１９×２６ドットの外接矩形７２を得た）。ここで、ｎ＝１９，ｍ＝２６、ｎ'＝３０，ｍ'＝３０であるから、ｎおよびｍがｎ'より小さくかつ、ｎおよびｍがｍ'より小さいので拡大正規化処理部１２３で文字サイズの拡大正規化を行ない、図８に示すように文字サイズが拡大され正規化された文字画像８１を得る。次に、正規化された文字画像８１に対し太め処理部１２５で太め処理を行ない、図９に示すような太さが３ドットで、大きさが３０×３０ドットの文字画像９１を得る。
【００２９】
なお、外接矩形検出部１２２で細線化された文字画像の外接矩形が正規化画像の所定サイズｎ'×ｍ'ドットより大きい場合（例えば、３５×２４ドットとか２１ドット×３４ドットとか、３６×４０ドットといった場合）には、ＯＲ正規化部１２４でサイズが縮小された正規化画像を得て、太め処理部１２５で太め処理を行ない、所定サイズｎ'×ｍ'ドット（本例では、３０×３０ドット）の文字画像９１を得る。
【００３０】
このようにして、正規化処理ブロック１２０で生成された正規化画像（図９）は線分の太さの変化の大きい部分（例えば、図６に示したような書き始めの太い部分や、途中でかすれている部分）もほぼ同一の線幅にされる。従って、文字の幅に対応した複数の辞書を要することなく、文字認識部１３０での安定した文字認識が実現する。
【００３１】
＜実施形態２＞
本発明の正規化方法の第２は、切出された文字画像の外接矩形を検出し、外接矩形の大きさにより拡大正規化または縮小正規化を行なった後、入力画像の線分の細線化処理を行なって得られた線分の幅をほぼ一定の太さにする。
【００３２】図２は、上記第２の正規化方法を適用した文字認識装置の一実施例の構成図であり、文字認識装置２０は文字切り出し部１１０、正規化処理ブロック２２０、および文字認識部１３０から構成され、正規化処理ブロック２２０は、外接矩形検出部２２１、拡大正規化部２２２、ＯＲ正規化部２２３、細線化処理部２２４、および太め処理部２２５を有している。
【００３３】
文字切り出し部１１０はスキャナー等の文字読取り装置で読み込んだ画像Ｉｍから一文字ずつ文字画像を切出す（図６）。
【００３４】
正規化処理ブロック２２０は、一文字分の画像の外接矩形を検出して外接矩形の大きさにより拡大正規化または縮小正規化を行なった後、切出された一文字分の画像の線分の細線化処理を行ない、得られた線分の幅を予め設定された太さにして認識処理部１３０に与える。そのために、先ず、文字切り出し部１１０によって切出された文字画像に対して外接矩形検出部２２１でその文字の外接矩形を検出し、外接矩形の大きさによって、その画像の拡大または縮小判定を行なう。なお、拡大／縮小判定の方法は実施形態１の外接矩形検出部１２１と同様でよい。
【００３５】
外接矩形検出部２２１で拡大判定がなされた場合には、拡大正規化部２２２で当該画像を拡大正規化処理して拡大し、サイズが正規化された文字画像を得る。また、縮小判定がなされた場合には、ＯＲ正規化部２２３で当該画像をＯＲ正規化処理してサイズが正規化された文字画像を得る。
【００３６】
上記拡大正規化部２２２またはＯＲ正規化部２２３でサイズが正規化された文字画像に対して細線化処理部２２１で細線化処理を行ない、その細線化画像を得る。そして、細線化処理部２２１で細線化された文字画像に対して太め処理部２２５で予め設定された太さになるようにする。
【００３７】
文字認識部１３０は、正規化処理ブロック２２０で生成された正規化画像をもとに文字の認識を行ない、その文字コードを認識結果として出力する。
【００３８】
これにより、実施形態１の場合と同様に、正規化処理ブロック２２０で生成された正規化画像（図９）は線分の太さの変化の大きい部分もほぼ同一の線幅にされる。従って、文字の幅に対応した複数の辞書を要することなく、文字認識部１３０での安定した文字認識が実現する。
【００３９】
＜実施形態３＞
本発明の正規化方法の第３は、切出された文字画像の外接矩形を検出し、文字画像の画素数が外接矩形に占める割合の大小によって、間引き正規化、拡大正規化または縮小正規化を行なった後、拡大正規化または縮小正規化を行なった文字画像については太め処理を施し、線分の幅をほぼ一定の太さにし、間引き正規化を行なった文字画像はそのまま出力する。なお、文字の線幅を一定にしなくても文字認識に悪影響を与えない場合、すなわち、外接矩形が所定値Ｃより大きい場合については太め処理を施すことなく、サイズの間引き処理を行なう。また、外接矩形の大きさが所定値ＤとＣの間にある場合についても太め処理を施すことなくサイズの拡大正規化処理を行なって処理速度を高める。
【００４０】
図３は、上記第３の正規化方法を適用した文字認識装置の一実施例の構成図であり、文字認識装置３０は文字切り出し部１１０、正規化処理ブロック３２０、および文字認識部１３０から構成され、正規化処理ブロック３２０は、外接矩形検出部３２１、文字画素検出部３２２、細線化処理部３２３、外接矩形検出部３２４、拡大正規化部３２５、ＯＲ正規化部３２６、間引き処理部３２７、および太め処理部３２８を有している。
【００４１】
文字切り出し部１１０はスキャナー等の文字読取り装置で読み込んだ画像Ｉｍから一文字ずつ文字画像を切り出す。
【００４２】
正規化処理ブロック３２０は、一文字分の画像の外接矩形を検出して、文字画像の画素数が外接矩形に占める割合の大小によって、間引き正規化、拡大正規化または縮小正規化を行なった後、拡大正規化または縮小正規化を行なった文字画像については太め処理を施し、得られた線分の幅を予め設定された太さにして認識処理部１３０に与える。なお、文字の線幅を一定にしなくても文字認識に悪影響を与えない場合、すなわち、外接矩形が所定値Ｃより大きい場合については太め処理を施すことなく、サイズの間引き処理を行ない、文字認識部１３０に出力する。また、外接矩形の大きさが所定値ＤとＣの間にある場合についても太め処理を施すことなくサイズの拡大正規化処理を行ない、文字認識部１３０に出力する。
【００４３】
そのために、外接矩形検出部３２１は、文字切り出し部３１０で切り出された１文字分の文字画像に対し文字の外接矩形を検出する。
【００４４】
文字画素検出部３２２は、文字切り出し部３１０で切り出した１文字分の文字画像の文字の画素数を検出すると共に、外接矩形検出部３２１で検出された外接矩形に対する文字画素の占める割合（以下、画素率Ｒ）を検出する。そして、外接矩形の大きさをＬとし、Ａ、Ｂ、Ｃ、Ｄを閾値とするとき、Ａ≦Ｒ≦Ｂで、かつＬ＞Ｃのとき、当該文字画素に対し間引き正規化のみを行なうと判定し、Ａ≦Ｒ≦Ｂで、かつＤ≦Ｌ≦Ｃのとき、当該文字画素に対し拡大正規化のみを行なうと判定し、Ａ≦Ｒ≦Ｂで、かつＬ＜Ｄのとき、または、Ｒ＜ＡまたはＲ＞Ｂのとき、当該文字画素に対し細線化処理を行なうと判定する。
【００４５】
細線化処理部３２３は、文字画素検出部３２２で細線化処理を行なうと判定された場合に、文字切り出し部１１０によって切出された１文字分の文字画像の細線化処理を行ない、その細線化画像を得る。次に、細線化処理部１２１で細線化された文字画像に対して外接矩形検出部３２４でその文字の外接矩形を検出し、外接矩形の大きさによって、その画像の拡大または縮小判定を行なう。なお、拡大／縮小判定の方法は実施形態１の外接矩形検出部１２１と同様でよい。
【００４６】
拡大正規化部３２５は、文字画素検出部３２２または外接矩形検出部３２４で拡大判定がなされた文字画像に対し、拡大正規化処理を施してサイズが正規化された画像を得る。
【００４７】
ＯＲ正規化部３２６は、外接矩形検出部３２４で縮小判定がなされた文字画像に対し、ＯＲ正規化処理を施してサイズが正規化された画像を得る。
【００４８】
間引き正規化部３２７は、文字画素検出部３２２で間引き判定がなされた文字画像に対し、間引き正規化処理を施してサイズが正規化された画像を得る。
【００４９】
太め処理部３２８では、拡大正規化部３２５またはＯＲ正規化部３２６でサイズが正規化された画像に対して画像の線分の幅が予め設定された太さになるように太め処理を施して文字認識部１３０に出力する。なお、文字画素検出部３２２で拡大正規化のみを行なうと判定された文字画像については拡大正規化部３２５での拡大正規化処理後、太め処理を施すことなく文字認識部１３０に出力する。また、同様に、文字画素検出部３２２で間引き処理を行なうと判定された文字画像については間引き正規化部３２７での拡大正規化処理後、太め処理を施すことなく文字認識部１３０に出力する。
【００５０】
文字認識部１３０は、正規化処理ブロック３２０から出力された正規化画像または間引き処理後の画像をもとに文字の認識を行ない、その文字コードを認識結果として出力する。
【００５１】
これにより、実施形態１の場合と同様に、正規化処理ブロック３２０から出力された正規化画像は線分の太さの変化の大きい部分もほぼ同一の線幅にされる。従って、文字の線幅に対応した複数の辞書を要することなく文字認識部１３０での安定した文字認識が実現する。
【００５２】
＜実施形態４＞
本発明の正規化方法の第４は、切出された文字画像の外接矩形を検出し、文字画像の画素数が外接矩形に占める割合の大小によって、拡大正規化、ＯＲ正規化または間引き正規化を行なった後、それら正規化画像に細線化処理等を施し、更に、太め処理を施して線分の幅をほぼ一定の太さにする。なお、文字の線幅を一定の太さにしなくても文字認識に悪影響を与えない場合、すなわち、外接矩形に対する文字の画素数の占める割合が所定値Ｅより小さく、かつ当該外接矩形の大きさが所定値Ｆより小さい場合については太め処理を施すことなく当該文字画像サイズの間引き処理を行なう。また、外接矩形に対する文字の画素数の占める割合が所定値ＧとＨの間にある場合についても太め処理を施すことなく、サイズの拡大正規化処理を行なって処理速度を高める。
【００５３】
図４は、上記第４の正規化方法を適用した文字認識装置の一実施例の構成図であり、文字認識装置４０は文字切り出し部１１０、正規化処理ブロック４２０、および文字認識部１３０から構成され、正規化処理ブロック４２０は、外接矩形検出部４２１、文字画素検出部４２２、拡大正規化部４２３、ＯＲ正規化部４２４、間引き処理部４２５、細線化処理部４２６および太め処理部４２７を有している。
【００５４】
文字切り出し部１１０はスキャナー等の文字読取り装置で読み込んだ画像Ｉｍから一文字ずつ文字画像を切り出す。
【００５５】
正規化処理ブロック４２０は、一文字分の画像の外接矩形を検出して、文字画像の画素数が外接矩形に占める割合の大小によって、拡大正規化、ＯＲ正規化または間引き正規化を行なった後、それら正規化画像に細線化処理等を施し、更に、太め処理を施して文字認識部１３０に出力する。なお、文字の線幅を一定の太さにしなくても文字認識に悪影響を与えない場合、すなわち、外接矩形に対する文字の画素数の占める割合が所定値Ｅより小さく、かつ当該外接矩形の大きさが所定値Ｆより小さい場合については太め処理を施すことなく当該文字画像サイズの間引き処理を行なう。また、外接矩形に対する文字の画素数の占める割合が所定値ＧとＨの間にある場合についても太め処理を施すことなく、サイズの拡大正規化処理を行なって文字認識部１３０に出力する。
【００５６】
そのために、先ず、文字切り出し部１１０によって切出された文字画像に対して外接矩形検出部４２１でその文字の外接矩形を検出し、外接矩形の大きさによって、その画像の拡大または縮小判定を行なう。なお、拡大／縮小判定の方法は実施形態１の外接矩形検出部１２１と同様でよい。
【００５７】
文字画素検出部４２２は、文字切り出し部１１０で切り出した１文字分の文字画像の文字の画素数を検出すると共に、外接矩形検出部４２１で検出された外接矩形に対する文字画素の占める割合（以下、画素率Ｒ）を検出する。そして、外接矩形の大きさをＬとし、Ｅ、Ｆ、Ｇ、Ｈを閾値とするとき、Ｒ≧Ｅ、または、外接矩形検出部４２１で拡大判定されていた場合でＬ＜Ｆのとき、当該文字画素に対し拡大正規化を行なった後に細線化処理を行なうものと判定し、Ｒ＜Ｅで、かつ外接矩形検出部４２１で拡大判定されていた場合で、Ｌ≧Ｆのとき、当該文字画素に対し拡大正規化のみを行なうと判定し、Ｒ＜Ｇで、かつ外接矩形検出部４２１で縮小判定されていた場合に、当該文字画素に対しＯＲ正規化を行なった後に細線化処理を行なうものと判定し、Ｒ＞Ｈで、かつ外接矩形検出部４２１で縮小判定されていた場合に、当該文字画素に対し間引き正規化を行なった後に細線化処理を行なうものと判定し、Ｇ≦Ｒ≦Ｈで、かつ外接矩形検出部４２１で縮小判定されていた場合に、当該文字画素に対し間引き正規化のみを行うものと判定する。
【００５８】
拡大正規化部４２３は、文字画素検出部４２２で拡大正規化の判定がなされた文字画像に対し、拡大正規化処理を施してサイズが正規化された画像を得る。
【００５９】
ＯＲ正規化部４２４は、文字画素検出部４２２でＯＲ正規化の判定がなされた文字画像に対し、ＯＲ正規化処理を施してサイズが正規化された画像を得る。
【００６０】
間引き正規化部４２５は、文字画素検出部４２２で間引き正規化判定がなされた文字画像に対し、間引き正規化処理を施してサイズが正規化された画像を得る。
【００６１】
細線化処理部４２６は、文字画素検出部４２２で細線化処理を行なうと判定された場合に、拡大正規化部４２３、ＯＲ正規化部４２４、または間引き正規化部でサイズの正規化が行なわれた文字画像に対し、細線化処理を行ない、その細線化画像を得る。太め処理部４２７では、細線化処理部４２６で得た細線化画像に対して画像の線分の幅が予め設定された太さになるように太め処理を施す。なお、文字画素検出部４２２で拡大正規化のみを行なうと判定された文字画像については拡大正規化部４２３での拡大正規化処理後、細線化処理及び太め処理を施すことなく文字認識部１３０に出力する。また、同様に、文字画素検出部４２２で間引き処理を行なうと判定された文字画像については間引き正規化部４２５での拡大正規化処理後、細線化処理および太め処理を施すことなく文字認識部１３０に出力する。
【００６２】
文字認識部１３０は、正規化処理ブロック４２０で生成された正規化画像をもとに文字の認識を行ない、その文字コードを認識結果として出力する。
【００６３】
これにより、実施形態１の場合と同様に、正規化処理ブロック４２０で生成された正規化画像は線分の太さの変化の大きい部分もほぼ同一の線幅にされる。従って、文字の線幅に対応した複数の辞書を要することなく文字認識部１３０での安定した文字認識が実現する。
【００６４】
【発明の効果】
本発明の文字画像正規化方法および文字認識装置によれば、文字画像の文字の線幅をほぼ一定にできるので、従来技術のように文字の線幅に対応した複数の辞書を要することなく、文字認識部での安定した文字認識が実現でき文字認識の信頼性が向上した。
【図面の簡単な説明】
【図１】第１の正規化方法を適用した文字認識装置の一実施例の構成図である。
【図２】第２の正規化方法を適用した文字認識装置の一実施例の構成図である。
【図３】第３の正規化方法を適用した文字認識装置の一実施例の構成図である。
【図４】第４の正規化方法を適用した文字認識装置の一実施例の構成図である。
【図５】読み込み画像Ｉｍの一例を示す図である。
【図６】図５の画像から切出された文字画像を示す図である。
【図７】細線化処理された文字画像の例を示す図である。
【図８】拡大正規化された文字画像の例を示す図である。
【図９】正規化された文字画像の例を示す図である。
【符号の説明】
６１入力画像
７１文字画像
８１拡大正規化された文字画像
９１所定の線幅の文字画像
１２１，２２４，３２３，４２６細線化処理手段
１２２，２２１，４２２外接矩形検出部
１２３，２２２，３２５，４２３拡大正規化部
１２４，２２３，３２６，４２４ＯＲ正規化部（縮小正規化部）
３２１外接矩形検出部（第１の外接矩形検出部）
３２４外接矩形検出部（第２の外接矩形検出部）
３２７，４２５間引き正規化部
３２８，４２７太め処理部[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a character recognition device, and more particularly, to a technique for normalizing character images having different writing line thicknesses and character sizes.
[0002]
[Prior art]
In the character recognition device, the line thickness of the read character image is not uniform, the character size is not constant, and character images having various line thicknesses and sizes can be obtained. In order to perform character recognition from such a character image, the read character image is normalized as preprocessing.
[0003]
As a method for normalizing the thickness and the size of the line of the read character image, a thinning normalization technique and an OR normalization technique are known. Here, thinning-out normalization is a method in which pixels are extracted at regular intervals and other pixels are truncated to generate pixels smaller in size than the input pixels. This is a method of generating an image having a smaller size than the input image by taking (logical sum).
[0004]
However, in the case of thinning normalization, there is a problem that fine lines are easily lost or faded, and there is a problem that OR normalization causes pixels in a character loop to be easily crushed. was there.
[0005]
In consideration of the above problems, there is a method of performing line segment thinning processing on an input image and performing OR normalization on the image after the thinning processing (Japanese Patent Laid-Open No. 3-37783).
[0006]
Further, as an improvement of the technique disclosed in the above-mentioned Japanese Patent Application Laid-Open No. 3-37783, a line segment thinning process is performed on an input image while changing its strength, and an OR operation is performed on each of the thinned character images. A method of obtaining a plurality of normalized character images for one image by performing normalization, performing recognition processing on each image, and obtaining a final recognition result from the obtained plurality of recognition results (JP-A-4-687). is there.
[0007]
[Problems to be solved by the invention]
In the character recognition technique disclosed in the above-mentioned Japanese Patent Application Laid-Open No. 4-687, since the strength of the line segment thinning process applied to the input image is changed to perform OR normalization on each of the thinned character images, one character In the case where the degree of change in the line segment thickness is large, the number of narrowing processes increases in accordance with the degree, the number of characters normalized by OR normalization also increases, and the number of recognition processes increases. As a result, when the change in the width of the line segment is more than a certain value, there is a certain limitation in terms of the memory capacity and the recognition processing speed, for example, the processing speed becomes slow.
[0008]
For example, if you take characters written with a ballpoint pen as an example, ink may accumulate on the pen tip of the ballpoint pen, and if you continue writing as it is, the line of entry will become thicker or thinner in one character The change in the thickness of the line segment in the character may be large. Also, in the case of a character enlarged by copying or the like, the line width becomes thicker when the magnification is large, and the image quality becomes coarse.
[0009]
There are various types of writing tools such as pencils, ballpoint pens, felt-tip pens, etc., and there are various types of core thicknesses. As described above, when a character written with a writing instrument and having a character having any line width is subjected to the same pattern matching recognition processing, a large number of dictionaries corresponding to the line width are required by using a feature extraction method in which feature data changes depending on the line width. As a result, there is a problem that the efficiency of the recognition processing is reduced.
[0010]
The present invention has been made in view of the above-described problems of the related art, and enables stable recognition without requiring a dictionary corresponding to the line width even when the thickness of a line in one character varies. An object of the present invention is to provide a character recognition apparatus including a character image normalization method and a normalization processing unit.
[0011]
[Means for Solving the Problems]
In order to solve the above-mentioned problem, a character recognition device that extracts a character image from an input image and performs a normalization process, and then performs a character recognition process on the normalized character image. A step of detecting a circumscribed rectangle, and, based on the size of the circumscribed rectangle and the ratio of the number of pixels of the character occupying the circumscribed rectangle, to the cut-out character image, Whether to perform only expansion normalization, only thinning normalization, or thinning And a normalization process for enlarging the character image based on the above determination result. Only thinning normalization processing or thinning processing Performing the thinning process, and further comprising, after the thinning process, a character image after the thinning process. Bounding rectangle And determining whether to enlarge or reduce the character image based on the size of the character image, and enlargement or reduction normalization processing of the thinned character image based on the determination result. Performing a thickening process on the character image subjected to the enlargement normalization process or the reduction normalization process to obtain a character image having a line width of a predetermined thickness. .
[0012]
A character image normalizing method according to a second aspect of the present invention is a character recognition device that extracts a character image from an input image, performs a normalization process, and then performs a character recognition process on the normalized character image. Detecting the circumscribed rectangle of the cut-out character image, and determining the size of the circumscribed rectangle and the ratio of the number of pixels of the character occupying the circumscribed rectangle to the cut-out character image. Whether to perform only expansion normalization, only thinning normalization, or to perform thinning after performing expansion normalization, thinning normalization, or reduction Determining, based on the determination result , A step of performing only the enlargement normalization processing of the character image or only the thinning normalization processing; Based on the above determination result, Character image Perform thinning processing after performing any of the expansion normalization processing, thinning normalization processing, or reduction processing Process and With this thinning process Thickening of thinned character images Give Producing a character image having a line width of a predetermined thickness.
[0013]
A character recognition device according to a third aspect of the present invention is a character recognition device that extracts a character image from an input image, performs a normalization process, and then performs a character recognition process on the normalized character image. A first circumscribing rectangle detecting means for detecting a circumscribing rectangle of the extracted character image, and a ratio between the size of the circumscribing rectangle and the number of pixels of the character occupying the circumscribing rectangle. Whether only expansion normalization, thinning normalization, or thinning Character pixel detection means for determining whether or not to perform, enlargement normalization processing of the character image based on the determination result of the character pixel detection means, or Enlargement normalizing means for performing enlargement normalization processing of a character image after thinning processing based on the determination result of the second circumscribed rectangle detection means; Based on the determination result of the character pixel detecting means, Decimation normalization processing Do Decimation normalization means And thinning processing means for performing thinning processing of the character image based on the determination result of the character pixel detecting means, and further detects a circumscribed circumscribed rectangle of the character image after the thinning processing and determines its size. A second circumscribing rectangle detecting means for determining whether to enlarge or reduce the character image by normalization; A reduction normalization unit that performs a reduction normalization process on the character image after the thinning process based on the determination result of the second circumscribed rectangle detection unit; Based on the determination result of the second circumscribed rectangle detecting means , The character image enlarged and normalized by the enlargement normalizing means or the character image reduced and normalized by the reduction normalization means Thickening means for applying a thickening process to a character image having a line width of a predetermined thickness. Equipped It is characterized by the following.
[0014]
A character recognition device according to a fourth aspect of the present invention is a character recognition device that extracts a character image from an input image, performs a normalization process, and then performs a character recognition process on the normalized character image. Circumscribing rectangle detection means for detecting a circumscribing rectangle of the extracted character image, and, based on the size of the circumscribing rectangle and the ratio of the number of pixels of the character occupying the circumscribing rectangle, to perform only enlargement normalization on the extracted character image, Do only thinning normalization, Specifies whether to perform thinning after performing expansion normalization, thinning normalization, or reduction. Character pixel detection means for determining, enlargement normalization means for performing enlargement normalization processing of the character image based on the determination result of the character pixel detection means, and thinning normalization of the character image based on the determination result of the character pixel detection means And a reduction normalization unit that performs a reduction normalization process on a character image based on the determination result of the character pixel detection unit, and further includes a determination result of the character pixel detection unit. The thinning is performed on the character image after performing one of the expansion normalization processing by the expansion normalization means, the thinning normalization processing by the thinning normalization processing means, or the reduction normalization processing by the reduction normalization means based on Do the processing Thinning means; The thinning processing is performed by the thinning processing means. Thickness processing for character images Giving Thickening processing means for forming a character image having a line width of a predetermined thickness; Equipped And features.
[0015]
[0016]
[0017]
[0018]
[0019]
BEST MODE FOR CARRYING OUT THE INVENTION
The character image normalization method of the present invention detects a circumscribed rectangle of a line segment of a cut-out character image, and performs normalization processing such as enlargement, reduction, thinning out, or thinning based on the size of the circumscribed rectangle. so Make the width of the line segment almost constant.
[0020]
<First embodiment>
In the first normalization method of the present invention, after performing a thinning process of a line segment of a cut-out character image, a circumscribed rectangle is detected, and expansion normalization or reduction normalization is performed based on the size of the circumscribed rectangle. The width of the obtained line segment is made constant by performing a thickening process on the obtained line segment.
[0021]
FIG. 1 is a configuration diagram of one embodiment of a character recognition device to which the first normalization method is applied. The character recognition device 10 includes a character cutout unit 110, a normalization processing block 120, and a character recognition unit 130. The normalization processing block 120 includes a thinning processing unit 121, a circumscribed rectangle detection unit 122, an expansion normalization unit 123, an OR normalization unit 124, and a thick processing unit 125.
[0022]
The character cutout unit 110 cuts out character images one by one from the image Im read by a character reading device such as a scanner.
[0023]
The normalization processing block 120 detects the circumscribed rectangle after performing the line-thinning processing of the line of the extracted image of one character, performs enlargement normalization or reduction normalization according to the size of the circumscribed rectangle, and performs size reduction. Is made constant, and the width of the obtained line segment is set to a predetermined thickness and given to the recognition processing unit 130. For this purpose, first, the thinning processing section 121 performs thinning processing of the character image of one character cut out by the character cutout section 110, and obtains the thinned image. Next, a circumscribed rectangle of the character is detected by the circumscribed rectangle detection unit 122 with respect to the image thinned by the thinning processing unit 121, and the enlargement or reduction of the character image is determined based on the size of the circumscribed rectangle. That is, in the circumscribed rectangle detection unit 122, when the size of the circumscribed rectangle is n × m dots and n ′ × m ′ is the normalized size of the character image, n and m are smaller than n ′, and n and m are smaller than n ′. When it is smaller than m ′, it is determined that the character image needs to be enlarged, and when at least one of n and m is larger than n ′ or n, it is determined that the character image needs to be reduced.
[0024]
When the circumscribed rectangle detection unit 122 determines the enlargement, the enlargement normalization unit 123 enlarges and normalizes the character image thinned by the thinning processing unit 122 to obtain an image whose size is normalized. When the reduction is determined, the character image thinned by the thinning processing unit 122 is OR-normalized by the OR normalizing unit 124 to obtain an image whose size is normalized.
[0025]
The thickening processing section 125 performs thickening processing on the character image whose size has been normalized by the enlargement normalizing section 123 or the OR normalizing section 124 so that the line segment width of the character image becomes a predetermined thickness. .
[0026]
The character recognition unit 130 performs character recognition based on the normalized image generated by the normalization processing block 120, and outputs the character code as a recognition result.
[0027]
[Example]
An example of the normalization block 120 according to the present embodiment will be described below. 5 shows an example of an input image, FIG. 6 shows an example of a character image cut out from the image of FIG. 5, and FIG. 7 shows an example of a character image subjected to thinning processing. 8 is a diagram illustrating an example of a character image that has been enlarged and normalized, and FIG. 9 is a diagram illustrating an example of a normalized character image. In a case where the image Im (FIG. 5) read from a scanner or the like is cut out one character at a time and normalized to a character image having a line width of 3 dots and a size of 30 × 30 dots (FIG. 9), In the case of normalizing the character image of "2" (FIG. 6) in which the writing start is thick and the middle is faint due to the accumulation of ink at the pen tip of the ballpoint pen, out of the character images of five characters cut out by 120 State.
[0028]
The character image 61 (FIG. 6) representing the character “2” captured by the normalization block 120 is first subjected to thinning processing by the thinning processing unit 121 to become a thinned character image as shown in FIG. 7. Next, a circumscribed rectangle (FIG. 7) of the character image 71 thinned by the circumscribed rectangle detection unit 122 is obtained (in the example of FIG. 7, a circumscribed rectangle 72 of 19 × 26 dots is obtained). Here, since n = 19, m = 26, n ′ = 30, and m ′ = 30, n and m are smaller than n ′, and n and m are smaller than m ′. The character size is enlarged and normalized, and a character image 81 whose character size is enlarged and normalized as shown in FIG. 8 is obtained. Next, thickening processing is performed on the normalized character image 81 by the thickening processing unit 125 to obtain a character image 91 having a thickness of 3 dots and a size of 30 × 30 dots as shown in FIG.
[0029]
When the circumscribed rectangle of the character image thinned by the circumscribed rectangle detection unit 122 is larger than a predetermined size of n ′ × m ′ dots of the normalized image (for example, 35 × 24 dots, 21 dots × 34 dots, or 36 × In the case of 40 dots, for example, a normalized image whose size is reduced by the OR normalizing unit 124 is obtained, the thickening processing unit 125 performs thickening processing, and a predetermined size of n ′ × m ′ dots (30 in this example). (× 30 dots) character image 91 is obtained.
[0030]
In this way, the normalized image (FIG. 9) generated by the normalization processing block 120 has a large change in line segment thickness (for example, a thick portion at the beginning of writing as shown in FIG. (Shaded portion) are also made to have substantially the same line width. Therefore, stable character recognition in the character recognition unit 130 is realized without requiring a plurality of dictionaries corresponding to character widths.
[0031]
<Embodiment 2>
A second method of the normalization method of the present invention is to detect a circumscribed rectangle of a cut-out character image, perform enlargement normalization or reduction normalization according to the size of the circumscribed rectangle, and then thin the line segment of the input image. The width of the line segment obtained by performing the processing is made substantially constant.
FIG. 2 is a block diagram of an embodiment of a character recognition device to which the second normalization method is applied. The character recognition device 20 includes a character cutout unit 110, a normalization processing block 220, and a character recognition unit. The normalization processing block 220 includes a circumscribed rectangle detection unit 221, an enlarged normalization unit 222, an OR normalization unit 223, a thin line processing unit 224, and a thick processing unit 225.
[0033]
The character cutout unit 110 cuts out character images one by one from the image Im read by a character reading device such as a scanner (FIG. 6).
[0034]
The normalization processing block 220 detects a circumscribed rectangle of the image of one character, performs enlargement normalization or reduction normalization according to the size of the circumscribed rectangle, and then thins out the line of the extracted image of one character. Processing is performed, and the width of the obtained line segment is set to a predetermined thickness and given to the recognition processing unit 130. For this purpose, first, a circumscribed rectangle of the character is detected by the circumscribed rectangle detection unit 221 from the character image cut out by the character cutout unit 110, and the enlargement or reduction of the image is determined based on the size of the circumscribed rectangle. . The enlargement / reduction determination method may be the same as that of the circumscribed rectangle detection unit 121 of the first embodiment.
[0035]
If the circumscribed rectangle detection unit 221 determines the enlargement, the enlargement normalization unit 222 enlarges the image by enlargement normalization processing and enlarges the image to obtain a character image whose size is normalized. When the reduction is determined, the image is OR-normalized by the OR normalizing unit 223 to obtain a character image whose size is normalized.
[0036]
The thinning processing unit 221 performs thinning processing on the character image whose size has been normalized by the enlargement normalizing unit 222 or the OR normalizing unit 223, and obtains the thinned image. Then, the thickening processing unit 225 sets the thickness of the character image thinned by the thinning processing unit 221 to a predetermined thickness.
[0037]
The character recognition unit 130 performs character recognition based on the normalized image generated by the normalization processing block 220, and outputs the character code as a recognition result.
[0038]
As a result, similarly to the first embodiment, in the normalized image (FIG. 9) generated in the normalization processing block 220, the portion where the change in the thickness of the line segment is large has almost the same line width. Therefore, stable character recognition in the character recognition unit 130 is realized without requiring a plurality of dictionaries corresponding to character widths.
[0039]
<Embodiment 3>
A third method of the normalization method of the present invention is to detect a circumscribed rectangle of a cut-out character image, and to perform thinning normalization, enlargement normalization, or reduction normalization depending on the ratio of the number of pixels of the character image to the circumscribed rectangle. Is performed, the character image subjected to the enlargement normalization or the reduction normalization is subjected to thickening processing, the width of the line segment is set to a substantially constant thickness, and the character image subjected to the thinning-out normalization is output as it is. Note that if the line width of the character is not fixed, the character recognition is not adversely affected, that is, if the circumscribed rectangle is larger than the predetermined value C, the size thinning process is performed without performing the thickening process. Also, when the size of the circumscribed rectangle is between the predetermined values D and C, the processing speed is increased by performing the size expansion normalization processing without performing the thickening processing.
[0040]
FIG. 3 is a configuration diagram of an embodiment of a character recognition device to which the third normalization method is applied. The character recognition device 30 includes a character cutout unit 110, a normalization processing block 320, and a character recognition unit 130. The normalization processing block 320 includes a circumscribed rectangle detection unit 321, a character pixel detection unit 322, a thinning processing unit 323, a circumscribed rectangle detection unit 324, an enlarged normalization unit 325, an OR normalization unit 326, a thinning processing unit 327, And a thick processing section 328.
[0041]
The character cutout unit 110 cuts out character images one by one from the image Im read by a character reading device such as a scanner.
[0042]
The normalization processing block 320 detects the circumscribed rectangle of the image for one character, and performs thinning normalization, enlargement normalization, or reduction normalization based on the ratio of the number of pixels of the character image to the circumscribed rectangle. The character image that has been subjected to the enlargement or reduction normalization is subjected to thickening processing, and the width of the obtained line segment is set to a predetermined thickness and given to the recognition processing unit 130. If the character line width is not fixed and the character recognition is not adversely affected, that is, if the circumscribed rectangle is larger than a predetermined value C, the size is thinned out without performing the thickening process, and the character recognition is performed. Output to the unit 130. In addition, even when the size of the circumscribed rectangle is between the predetermined values D and C, the enlargement normalization process is performed without performing the thickening process, and the result is output to the character recognition unit 130.
[0043]
Therefore, the circumscribed rectangle detection unit 321 detects a circumscribed rectangle of a character in the character image of one character cut out by the character cutout unit 310.
[0044]
The character pixel detection unit 322 detects the number of pixels of a character in a character image of one character cut out by the character cutout unit 310, and also occupies the ratio of character pixels to the circumscribed rectangle detected by the circumscribed rectangle detection unit 321 (hereinafter, referred to as The pixel ratio R) is detected. When the size of the circumscribed rectangle is L and A, B, C, and D are threshold values, when A ≦ R ≦ B and L> C, only thinning normalization is performed on the character pixel. When A ≦ R ≦ B and D ≦ L ≦ C, it is determined that only the enlargement normalization is performed on the character pixel. When A ≦ R ≦ B and L <D, or When R <A or R> B, it is determined that thinning processing is to be performed on the character pixel.
[0045]
When the character pixel detection unit 322 determines that the thinning processing is to be performed, the thinning processing unit 323 performs the thinning processing of the character image of one character cut out by the character cutout unit 110, and performs the thinning. Get an image. Next, a circumscribed rectangle of the character is detected by the circumscribed rectangle detection unit 324 with respect to the character image thinned by the line thinning processing unit 121, and the enlargement or reduction of the image is determined based on the size of the circumscribed rectangle. The enlargement / reduction determination method may be the same as that of the circumscribed rectangle detection unit 121 of the first embodiment.
[0046]
The enlargement normalization unit 325 performs an enlargement normalization process on the character image for which enlargement has been determined by the character pixel detection unit 322 or the circumscribed rectangle detection unit 324, and obtains an image whose size has been normalized.
[0047]
The OR normalization unit 326 performs an OR normalization process on the character image for which the circumscribing rectangle detection unit 324 has performed the reduction determination, and obtains an image whose size has been normalized.
[0048]
The thinning-out normalizing section 327 performs a thinning-out normalizing process on the character image for which the character pixel detecting section 322 has made the thinning-out determination, and obtains an image whose size is normalized.
[0049]
The thickening processing unit 328 performs a thickening process on the image whose size has been normalized by the enlargement normalizing unit 325 or the OR normalizing unit 326 so that the width of the image line segment becomes a preset thickness. Output to the character recognition unit 130. Note that the character image that has been determined by the character pixel detection unit 322 to perform only the enlargement normalization is output to the character recognition unit 130 without performing the thickening process after the enlargement normalization process performed by the enlargement normalization unit 325. Similarly, the character image determined to be subjected to the thinning process by the character pixel detecting unit 322 is output to the character recognizing unit 130 without performing the thickening process after the enlargement normalizing process by the thinning normalizing unit 327.
[0050]
The character recognition unit 130 performs character recognition based on the normalized image output from the normalization processing block 320 or the image after the thinning processing, and outputs the character code as a recognition result.
[0051]
As a result, similarly to the first embodiment, the normalized image output from the normalization processing block 320 has substantially the same line width even in a portion where the line segment has a large change in thickness. Therefore, stable character recognition in the character recognition unit 130 is realized without requiring a plurality of dictionaries corresponding to the line width of characters.
[0052]
<Embodiment 4>
A fourth aspect of the normalization method of the present invention is to detect a circumscribed rectangle of a cut-out character image, and to perform enlargement normalization, OR normalization, or thinning normalization depending on the ratio of the number of pixels of the character image to the circumscribed rectangle. Is performed, the normalized images are subjected to a thinning process or the like, and further a thickening process is performed to make the width of the line segment almost constant. If the line width of the character does not have a certain thickness and does not adversely affect character recognition, that is, the ratio of the number of pixels of the character to the circumscribed rectangle is smaller than the predetermined value E, and the size of the circumscribed rectangle is Is smaller than the predetermined value F, the thinning process of the character image size is performed without performing the thickening process. Even when the ratio of the number of pixels of the character to the circumscribed rectangle is between the predetermined values G and H, the processing speed is increased by performing the size expansion normalization processing without performing the thickening processing.
[0053]
FIG. 4 is a configuration diagram of an embodiment of a character recognition device to which the fourth normalization method is applied. The character recognition device 40 includes a character cutout unit 110, a normalization processing block 420, and a character recognition unit 130. The normalization processing block 420 includes a circumscribed rectangle detection unit 421, a character pixel detection unit 422, an expansion normalization unit 423, an OR normalization unit 424, a thinning processing unit 425, a thinning processing unit 426, and a thick processing unit 427. are doing.
[0054]
The character cutout unit 110 cuts out character images one by one from the image Im read by a character reading device such as a scanner.
[0055]
The normalization processing block 420 detects the circumscribed rectangle of the image for one character, and performs enlargement normalization, OR normalization, or thinning normalization according to the ratio of the number of pixels of the character image to the circumscribed rectangle. The normalized images are subjected to thinning processing and the like, and further subjected to thickening processing and output to the character recognition unit 130. If the line width of the character does not have a certain thickness and does not adversely affect character recognition, that is, the ratio of the number of pixels of the character to the circumscribed rectangle is smaller than the predetermined value E, and the size of the circumscribed rectangle is Is smaller than the predetermined value F, the thinning process of the character image size is performed without performing the thickening process. In addition, even when the ratio of the number of pixels of the character to the circumscribed rectangle is between the predetermined values G and H, the enlargement normalization process is performed and the character recognition unit 130 is output without performing the thickening process.
[0056]
For this purpose, first, the circumscribed rectangle detection unit 421 detects the circumscribed rectangle of the character from the character image cut out by the character cutout unit 110, and determines whether to enlarge or reduce the image based on the size of the circumscribed rectangle. . The enlargement / reduction determination method may be the same as that of the circumscribed rectangle detection unit 121 of the first embodiment.
[0057]
The character pixel detection unit 422 detects the number of pixels of a character in a character image of one character cut out by the character cutout unit 110, and determines a ratio of a character pixel to a circumscribed rectangle detected by the circumscribed rectangle detection unit 421 (hereinafter, referred to as a character pixel). The pixel ratio R) is detected. When the size of the circumscribed rectangle is L and E, F, G, and H are threshold values, R ≧ E, or when L <F in the case where enlargement is determined by the circumscribed rectangle detection unit 421, When it is determined that thinning processing is to be performed after performing enlargement normalization on a character pixel, and R <E and the circumscribing rectangle detection unit 421 determines that the character pixel has been enlarged, and L ≧ F, the character pixel Is determined to perform only enlargement normalization, and if R <G and the circumscribed rectangle detection unit 421 determines reduction, the thinning processing is performed after performing the OR normalization on the character pixel. If R> H and the circumscribing rectangle detection unit 421 determines that the character pixel has been reduced, it is determined that thinning processing is performed after thinning normalization is performed on the character pixel, and G ≦ R ≦ H and the circumscribed rectangle detection unit 421 determines When it had determines that performs only thinning normalized to the character pixel.
[0058]
The enlargement normalization unit 423 performs enlargement normalization processing on the character image for which enlargement normalization has been determined by the character pixel detection unit 422 to obtain an image whose size has been normalized.
[0059]
The OR normalization unit 424 performs an OR normalization process on the character image for which the determination of the OR normalization has been performed by the character pixel detection unit 422 to obtain an image whose size has been normalized.
[0060]
The thinning-out normalizing unit 425 performs a thinning-out normalization process on the character image for which the thinning-out normalization is determined by the character pixel detecting unit 422 to obtain an image whose size is normalized.
[0061]
When the character pixel detection unit 422 determines that the thinning processing is to be performed, the thinning processing unit 426 performs size normalization by the enlargement normalization unit 423, the OR normalization unit 424, or the thinning normalization unit. A thinning process is performed on the character image thus obtained to obtain the thinned image. The thickening processing unit 427 performs thickening processing on the thinned image obtained by the thinning processing unit 426 so that the width of the line segment of the image becomes a predetermined thickness. Note that the character image that has been determined by the character pixel detection unit 422 to perform only enlargement normalization is processed by the character recognition unit 130 without performing thinning processing and thickening processing after the enlargement normalization processing performed by the enlargement normalization unit 423. Output. Similarly, for the character image determined to be subjected to the thinning process by the character pixel detecting unit 422, the character recognizing unit 130 does not perform the thinning process and the thickening process after the enlargement normalizing process in the thinning normalizing unit 425. Output to
[0062]
The character recognition unit 130 performs character recognition based on the normalized image generated by the normalization processing block 420, and outputs the character code as a recognition result.
[0063]
As a result, similarly to the first embodiment, in the normalized image generated in the normalization processing block 420, the portion where the line segment has a large change in thickness is made to have substantially the same line width. Therefore, stable character recognition in the character recognition unit 130 is realized without requiring a plurality of dictionaries corresponding to the line width of characters.
[0064]
【The invention's effect】
According to the character image normalization method and the character recognition device of the present invention, the line width of the characters in the character image can be made substantially constant, so that a plurality of dictionaries corresponding to the line widths of the characters are not required as in the related art. Stable character recognition in the character recognition unit was realized, and the reliability of character recognition was improved.
[Brief description of the drawings]
FIG. 1 is a configuration diagram of an embodiment of a character recognition device to which a first normalization method is applied.
FIG. 2 is a configuration diagram of an embodiment of a character recognition device to which a second normalization method is applied.
FIG. 3 is a configuration diagram of an embodiment of a character recognition device to which a third normalization method is applied.
FIG. 4 is a configuration diagram of an embodiment of a character recognition device to which a fourth normalization method is applied.
FIG. 5 is a diagram illustrating an example of a read image Im.
FIG. 6 is a view showing a character image cut out from the image of FIG. 5;
FIG. 7 is a diagram illustrating an example of a character image subjected to a thinning process.
FIG. 8 is a diagram illustrating an example of an enlarged and normalized character image.
FIG. 9 is a diagram illustrating an example of a normalized character image.
[Explanation of symbols]
61 Input image
71 character image
81 Enlarged Normalized Character Image
91 Character image with predetermined line width
121, 224, 323, 426 Thinning processing means
122,221,422 circumscribed rectangle detector
123, 222, 325, 423 Expansion normalization unit
124, 223, 326, 424 OR normalization unit (reduction normalization unit)
321 circumscribed rectangle detector (first circumscribed rectangle detector)
324 circumscribed rectangle detection unit (second circumscribed rectangle detection unit)
327,425 Thinning-out normalization part
328, 427 Thick processing unit

Claims

After extracting a character image from an input image and performing normalization processing, a character recognition device that performs character recognition processing on the normalized character image,
Detecting a circumscribed rectangle of the extracted character image;
From the size of the circumscribed rectangle and the ratio of the number of pixels of characters occupying the circumscribed rectangle, it is determined whether to perform only enlargement normalization, only thinning normalization, or thinning on the cut-out character image. The process of
Based on the determination result, only performing the enlargement normalization process of the character image, only the thinning normalization process, or performing a thinning process,
Further, in the subsequent stage of the thinning process, a step of detecting a circumscribed rectangle of the character image after the thinning process and determining whether to enlarge or normalize the character image based on its size, Performing an enlargement normalization process or a reduction normalization process on the thinned character image based on the determination result, and performing a fattening process on the enlargement normalization process or the reduction normalization process. Forming a character image having a line width of a predetermined thickness.

After extracting the character image from the input image and performing the normalization process, in a character recognition device that performs character recognition processing on the normalized character image, detecting a circumscribed rectangle of the extracted character image; From the size of the circumscribed rectangle and the ratio of the number of pixels of the character occupying the circumscribed rectangle, the cut-out character image is subjected to only enlargement normalization, only thinning normalization, or enlargement normalization, A step of determining whether to perform thinning after performing either thinning normalization or reduction ,
A step based on the determination result, for performing only the enlargement normalization processing of the character image only, or thinning normalization process,
Based on the determination result, and performing thinning processing after performing any of the expansion normalization processing of the character image, the thinning normalization or reduction processing, thinning character by the thinning process A step of performing a thickening process on the image to obtain a character image having a line width of a predetermined thickness;
A method for normalizing a character image, comprising:

After extracting a character image from an input image and performing normalization processing, a character recognition device that performs character recognition processing on the normalized character image,
First circumscribing rectangle detection means for detecting a circumscribing rectangle of the cut-out character image;
From the size of the circumscribed rectangle and the ratio of the number of pixels of characters occupying the circumscribed rectangle, it is determined whether to perform only enlargement normalization, only thinning normalization, or thinning on the cut-out character image. Character pixel detecting means,
Enlargement normalizing means for performing the enlargement normalization processing of the character image based on the judgment result of the character pixel detection means or the enlargement normalization processing of the character image after the thinning processing based on the judgment result of the second circumscribed rectangle detection means. When,
Thinning- out normalizing means for performing thinning- out normalizing processing of the character image based on the determination result of the character pixel detecting means,
Thinning processing means for performing thinning processing of the character image based on the determination result of the character pixel detecting means,
Further, a second circumscribed rectangle detecting means for detecting a circumscribed circumscribed rectangle of the character image after the thinning process and determining whether to enlarge or normalize the character image based on its size,
Reduction normalization means for performing reduction normalization processing of the character image after the thinning processing based on the determination result of the second circumscribed rectangle detection means;
On the basis of the determination result of the second circumscribing rectangle detection unit, a fattening process is performed on the character image enlarged and normalized by the enlargement normalization unit or the character image reduced and normalized by the reduction normalization unit. Thickening processing means for making a character image having a line width of a predetermined thickness,
Character recognition device, wherein a has, with a character normalization means.

After extracting a character image from an input image and performing normalization processing, a character recognition device that performs character recognition processing on the normalized character image,
Circumscribed rectangle detection means for detecting a circumscribed rectangle of the cut-out character image,
From the size of the circumscribed rectangle and the ratio of the number of pixels of the character occupying the circumscribed rectangle, the cut-out character image is subjected to only enlargement normalization, only thinning normalization, or enlargement normalization, Character pixel detection means for determining whether to perform thinning after performing either thinning normalization or reduction ,
Enlargement normalization means for performing enlargement normalization processing of the character image based on the determination result of the character pixel detection means,
Thinning-out normalizing means for performing thinning-out normalizing processing of the character image based on the determination result of the character pixel detecting means,
Reduction normalization means for performing reduction normalization processing of the character image based on the determination result of the character pixel detection means,
Further, based on the determination result of the character pixel detection unit , any one of an expansion normalization process by the expansion normalization unit, a thinning normalization process by the thinning normalization processing unit, or a reduction normalization process by the reduction normalization unit Thinning processing means for performing thinning processing on the character image after performing
Thickening processing means for performing a thickening process on the character image subjected to the thinning processing by the thinning processing means to obtain a character image having a line width of a predetermined thickness;
The a character recognition device which is characterized that you with a character normalization means.