JP3785306B2 - Image processing method and apparatus - Google Patents

Description

次に、図５は、Ｇプレーンの補完処理を示し、画素Ｇの周囲の４画素Ｇ１〜Ｇ４は、以下の式により作成する。
【００３０】

更に、図６は、Ｂプレーンの補完処理を示し、画素Ｂの周囲の８画素Ｂ１〜Ｂ８は、以下の式により作成する。
【００３１】

次に、Ｙプレーン画像（２４０ｄｐｉ）の生成を図７により、説明する。Ｙ画像は、モノクロ画像と称されている。各ＲＧＢ画像がそれぞれ固有のフィルタを経由して読まれた画像であるため、特定の色成分をドロップアウトしているのに対し、Ｙ画像は、各色の成分をそのまま出力する輝度データである。このＹプレーン画像は、ＲＧＢプレーン画像から生成できる。即ち、以下の比率計算式により、ＲＧＢからＹを生成する。
【００３２】
Ｙ＝０．２９９×Ｒ＋０．５８７×Ｇ＋０．１１４×Ｂ
上記式から、明らかなように、Ｇプレーン画像の比率が高いのは、Ｇプレーン内に色成分が多く含まれるためである。このように、各ＲＧＢを個々に分離し、結合することにより、モノクロ画像を生成できる。
【００３３】
図３に戻り、２値化処理部５１は、多値のＹ，Ｒ，Ｇ画像を２値化し、Ｙ，Ｒ，Ｇ画像（２値化）を生成する。前述のＹ，Ｒ，Ｇプレーン画像は、２５６階調を保有した多値イメージデータで構成されている。文字認識や印鑑照合等のコンピュータで認識、照合処理する場合は、性能等を考慮する必要があり、２値化処理部５１は、２５６階調の多値イメージデータを「０又は１」の２値化イメージデータに変換する。
【００３４】
ＯＣＲ画像生成部５２は、後述するように、以下の手順で、ＯＣＲ画像データを生成する。先ず、それぞれのプレーン画像はそれぞれの受光センサでフィルタされているので、赤ドロップアウトカラーであればＲプレーンを使用する。Ｒプレーン上では、ドロップアウトカラー部分はドロップアウトしており文字部分は低解像度（１２０ｄｐｉ相当）となっている。このＲプレーン画像を文字部分の情報欠落がないようにＯＲ間引き演算にて、１／４の大きさに縮小を行う。縮小されたプレーン画像中の背景部と文字部分でダイナミックな２階調分離（大津法２値化処理）を行う。大津法２値化されたプレーン画像を縮小前の大きさに画像を戻す。文字部分を潰れぎみにすることで後のＹ信号（輝度信号）とのフィルタで高精細な画像が取得できる。
【００３５】
文字部分を高精細にするためにＹ信号（輝度信号）を使用する。Ｙ信号は、全てのプレーン画像の演算合成であるので、もともと高精細（２４０ｄｐｉ相当）である。このＹ信号に対しては浮動的２値化処理（Niblack法２値化）を行い、なるべく情報（文字カスレ）を出すように２値化を行う。この場合、ドロップアウトカラー部分は浮動的２値化処理で黒画素（ドロップアウトしない）となるが、後でプレーン画像とフィルタ合成することで不必要な情報は消すことができる。Ｒプレーン２値とＹ信号（輝度信号）２値をフィルタすることで高精細なドロップアウトカラー画像を生成する。この場合、単純なＡＮＤマスク演算で画像生成が可能である。
【００３６】
尚、大津法２値化処理については、文献、大津展之："判別および最小２乗基準に基づく自動しきい値選定法", 電子情報通信学会論文誌, J63-D, vol.4, pp.349-356(1980)により詳細に開示されている。又、Ｎｉｂｌａｃｋ法２値化処理については、文献、W.Niblack:"An Introduction to Digital Image Processing",pp.115-116,Englewood Cliffs, N.J.:Prentice Hall(1986)により、詳細に示されている。
【００３７】
図８に沿い、帳票イメージ処理を説明する。
（Ｓ１００）上位アプリケーション３１は、制御部３０に読み取りを指示する。読み取り指示は、読み取り指示と、要求画像で構成されている。要求画像は、文字認識画像２値、印影画像２値のように指定される。勿論、この読み取り指示は、イメージリーダ１のベッド１０に帳票を搭載した後である。
（Ｓ１１０）制御部３０は、イメージリーダ１の読み取り回路１２にカラーＣＣＤ１１の読み取りを指示する。これにより、カラーＣＣＤ１１は、読み取りを開始し、アンプ１３、Ａ／Ｄ変換器１４を介しＲＧＢデータを出力する。
（Ｓ１２０）このＲＧＢデータは、画像メモリ２２に格納される。これにより、読み取りは終了し、オペレータは、帳票を取り外すことができる。
（Ｓ１３０）この格納後、制御部３０の指示により、画像処理部２１は、上位の要求した画像を生成する前述の画像処理を行う。これにより、画像処理部２１の図示しないメモリに各プレーン画像が格納される。
（Ｓ１４０）制御部３０は、上位の要求画像をプレーンメモリから上位アプリケーション３１に通知する。ＯＣＲ画像は、ＯＣＲエンジン３２で文字認識処理される。
【００３８】
このように、高解像度の輝度データを用いて、ドロップアウトカラーのＯＣＲ用画像データを生成できる。ドロップアウトカラーのため、文字枠等が除去され、高解像度の文字画像が得られる。従って、文字認識が正確となり、特に手書き漢字のような複雑な文字認識を正確にできる。
［ＯＣＲ画像生成］
以下、実施例として２４０ｄｐｉ赤ドロップアウトカラー画像作成の例で説明する。
【００３９】
図９は、ＯＣＲ画像処理のフロー図、図１０は、Ｒ（２値）データ生成動作の説明図、図１１は、大津法２値化処理の説明図、図１２は、輝度信号の２値化処理のフロー図、図１３は、そのＮｉｂｌａｃｋ２値化処理の説明図、図１４は、その文字のカスレ部分の推定処理の説明図、図１５は、そのフィルタ処理の説明図、図１６は、フィルタ処理の動作説明図である。
【００４０】
図９に示すように、原画像を入力して画素補完処理（プレーン毎に分割，輝度信号の作成）を行っても、Ｒプレーン５９は情報量が少ないため画素数では２４０ｄｐｉであるが解像度的には低解像度（１２０ｄｐｉ相当）である。
【００４１】
Ｒプレーン画像５９に対して１／４相当にＯＲ演算間引き６０を行う。ＯＲ演算間引きをする目的は、後の輝度信号画像とフィルタマスクする際に、輝度信号の情報（文字）を欠落させないためと、高速にＲプレーン５９を２値化するためである。図１０に示すように、１／４プレーンの画素Ｒａは、Ｒプレーンの画素Ｒ１〜Ｒ４の矩形マトリクス中で２番目に濃度が濃いものをセレクトする。
【００４２】
次に、１／４縮尺されたＲプレーン画像に対して大津法２値化処理６１を行う。Ｒプレーン画像は、ドロップアウトしているので、背景部と文字部とのダイナミックスライス（閾値）で２値化を行う。大津法２値化は、図１１に示すように、イメージ中のヒストグラム（各濃度値の画素数）を計算し、ヒストグラム（濃度分布）から、以下の式で、生起確率分布と全体濃度平均を求める。生起確率分布からクラス間分散を求めて２値化閾値を算出する。

【００４３】
このクラス間分散値が最大である濃度値を２値化スライスレベルとして決定し、各画素の濃度値をこのスライスレベルでスライスして、２値化する。
更に、２値化されたＲプレーン画像を元の大きさに戻す。即ち、４倍拡大処理６２して、２４０ｄｐｉ輝度信号とフィルタマスクするために元の大きさ２４０ｄｐｉの画像データ６３に戻す。
【００４４】
一方、図９及び図１２に示すように、輝度信号Ｙ多値６４に対してＮｉｂｌａｃｋ法２値化６５を行う。輝度信号は、２４０ｄｐｉの高精細画像であるため、浮動的に２値化をおこない情報をすべて出力する。但し、Ｎｉｂｌａｃｋ法２値化は、ソフトウェア上で処理するには時間がかかる処理であるので、文字のカスレ部分のみを推定（図１４）して、その部分のみＮｉｂｌａｃｋ法２値化を行う。
【００４５】
図１３に示すように、Ｎｉｂｌａｃｋ法２値化は、矩形フィルタ（フィルタの大きさは任意）を設定し、フィルタ内の濃度分散値ｇ２から背景部か情報部を判別する。濃度分散値ｇ２が分散閾値（σmin）以上であれば，注目画素［ｋ］の濃度の閾値Ｓを求め、注目画素の濃度値と閾値Ｓとを比較して、その白画素か黒画素かを判定する。局所領域に対して閾値を設定していくので、情報の欠落が少ない。
【００４６】
今回の実施例では、ＯＣＲ向け手書き文字認識画像の生成しているので７×７のフィルタを設定している（手書き文字のストローク幅は３ドット程度であるので７×７のフィルタが望ましい）。
【００４７】
又、 閾値を求めたい注目画素（ｋ）に対して、上下左右３画素ずつ矩形領域を設定するため、注目画素が背景（フィルタ内も背景のみ）であれば、図１３の背景閾値σｍｉｎで、判定式（１）により、背景と判断される。また、濃度絶対値を見ていないので、７×７フィルタ内に文字ストロークがすべて被ってしまった場合にも、背景と見なされてしまうが、後で述べる「文字カスレの推定」で文字ストローク部分内は黒に補完できる。フィルタ内に文字ストロークと背景が入っている場合には、濃度分散値ｇ２が、背景閾値σmin以上となり、図１３の注目画素ｋの閾値Ｓを図１３の式（２）から求める。注目画素が求まった閾値Ｓ以下の濃度であれば黒画素となり、以上であれば白画素と判断する。
【００４８】
更に、求まった図１３の（２）式のｒｏｏｔ（g²）に対して係数(-0.1)を乗算する。閾値Ｓをアンダーぎみにして，文字周辺のノイズを抑えるようにすることもできる。
【００４９】
次に、輝度信号の文字のカスレ部分を推定するために、Ｒプレーン２値化で使用した大津法２値化閾値算出方法を摘要する。図１４に示すように、輝度信号全体画像を大津法でダイナミックな２階調分離閾値Ｏｔを求め、その閾値Ｏｔを中心として低い濃度方向（黒となる濃度）と高い濃度方向（白となる濃度）に対して比率ａ，ｂを設定する。 即ち、文字カスレ部分推定エリアの算出は、下記の式による。
【００５０】
カスレ範囲の係数：a= 0.10 〜 0.90 b= 0.10 〜 0.90
黒領域の濃度平均：b#ave= 濃度分布0 〜 Ｏtの濃度平均
白領域の濃度平均：w#ave= 濃度分布Ｏt 〜 255の濃度平均
かすれ範囲の下限bt= （（1.0-a）× ot）＋（a×b#ave）
かすれ範囲の上限wt= （（1.0-b）× ot）＋（b×w#ave）
上記の推定エリア（ｗｔ〜ｂｔ）の濃度値の注目画素に対して、Ｎｉｂｌａｃｋ２値化を行い、それ以外のb#aveのエリアの画素は黒，w#aveエリアの画素は白と判断することにより、大幅な高速処理が可能である。ａ,ｂの比率は大きくすれば局所２値化の部分が多くなるが、あまり大きくしてしまうと手書き文字を２値化する場合には無駄な処理になってしまう。今回の実施例では、カスレ範囲の係数をa=0.35，b=0.62（全体の５〜８％をＮｉｂｌａｃｋ処理：文字記入数により比率は変動する）にすることで、１００パーセント処理を行った画像と同等レベルの結果が得られた。
【００５１】
そして、図９及び図１５、図１６に示すように、輝度信号２値データ６６とＲプレーン２値データ６３に対してマスクフィルタ合成処理６７する。即ち、それぞれ２値化された画像を単純ＡＮＤマスク合成することで、図１６の合成画像６８に示すように、ドロップアウト部分はＲプレーンの画像６３から、文字部分（情報部分）は輝度信号の画像６６から情報を取り出すことができる。
【００５２】
このように、各ＲＧＢの輝度であり、解像度が高い輝度データを使用する。しかし、輝度データは、特定色をドロップアウトできない。そこで、カラー画像データからドロップアウトカラー色の画像データを抽出し、この画像データで、高解像度の輝度データをフィルタリングすることにより、輝度データからドロップアウトカラーの部分を除去する。これにより、ドロップアウトカラーを除去した高解像度の輝度データをＯＣＲ用画像データとして、生成できる。これにより、高精度の文字認識が可能となる。
【００５３】
又、カラー画像データから２値化された輝度データを生成し、カラー画像データからドロップアウトした色の画像データを抽出し、補完処理して、色のプレーンデータを作成した後、プレーンデータを２値化処理して、色の画像データを作成することにより、２値データのフィルタリング処理が可能となり、フィルタリング処理が容易且つ高速となる。
【００５４】
更に、多値の前記カラー画像データから多値の輝度データを生成し、多値の輝度データに対し、設定された矩形フィルタ内の濃度分散値から２値化処理するＮｉｂｌａｃｋ法により、局所的に最適な２値化ができ、情報部分である文字の欠落を防止できる。
【００５５】
更に、プレーンデータの濃度平均に基ずいて、２値化閾値を算出し、プレーンデータを２値化閾値で２値化する大津法等の２値化処理により、輝度データの文字部分の欠落のないマスクデータを生成できる。
【００５６】
更に、２値化閾値を算出するステップは、プレーンデータのプレーンを縮小するステップと、縮小されたプレーンデータの濃度平均に基ずいて、２値化閾値を算出するステップからなり、２値化するステップは、縮小されたプレーンデータを２値化閾値で２値化するステップと、２値化されたプレーンデータのプレーンを拡大するステップとを有することにより、高速にマスクデータを生成できる。
【００５７】
実施例として、図２０の前述のＯＣＲ帳票の一例としての振込み依頼書４の画像データで、図１７及び図１８により、説明する。このＯＣＲ帳票４では、記入文字の周囲は、赤色枠４０で囲われている。又、押印４１は、ご依頼人の右側個所の「検印欄」になされている。ここでの印影は、システム処理する上で、銀行支店の担当者の検印である。押印４１も赤色である。
【００５８】
図１７に示すように、前述のＯＣＲ帳票４のＹ信号２値データ６６は、記入枠４０、記入文字、押印４１が現れている。一方、Ｒプレーン２値画像データ６３は、赤色の記入枠４０、押印４１がドロップアウトされ、記入文字が現れている。Ｙ信号２値データ６６を、Ｒプレーン２値データ６３でマスク合成処理６７を行い、ＯＣＲ用画像データ６８を得る。このＯＣＲ用画像データ６８の拡大図が図１８である。即ち、高解像度のＹ信号から、ドロップアウト色の部分が除去されていることがわかる。
【００５９】
このＯＣＲ用画像データの特徴を、図１９により、説明する。図１９は、記入枠に囲まれた「北海道拓殖」の手書き文字の、Ｙ信号２値画像６６を、Ｒプレーン２値画像６３、ＯＣＲ用画像６８を示す。Ｙ信号２値画像６６は、高解像度であるが、記入枠が残る。Ｒプレーン画像６３は、記入枠は除去されるが、文字は低解像度である。ＯＣＲ用画像６８は、高解像度であり、記入枠も除去されている。
【００６０】
前述の図１のシステムでは、上位アプリケーション３１は、読み取り指示に順序読みモードを指定し、要求画像に、１回目は、文字認識画像２値を、２回目は、印影画像２値を指定する。
【００６１】
制御部３０は、要求「文字認識画像２値」に対し、ＯＣＲ処理部５２で２値化されたＯＣＲ画像（２値）を上位に転送する。このＯＣＲ画像（２値）は、赤色系をドロップアウトするので、ＯＣＲ帳票の記入枠等の赤色系を落し、黒文字のみが残る。このため、文字認識画像として、最適な画像である。前述のように、ＯＣＲ画像（２値）は、高解像度の文字画像を表し、且つ文字認識の記入枠４０が全てドロップアウトされ、且つ押印４１もドロップアウトされている。
【００６２】
続いて、「印影画像２値」の画像要求に対し、２値化処理部５１で２値化されたＧ画像（２値）を上位に転送する。このＧ画像（２値）は、印影朱肉の赤色系を最も感度良く画像として表現する特徴を持つ。このため、印影画像として、最適な画像である。Ｇ画像（２値）は、印影４１が明確に画像として現れている。勿論、この時は、赤色系を画像出力するため、文字認識の記入枠４０等も画像に残る。上位３１は、Ｇ画像から、印影領域を切り出し、ホストへ転送する。ホストでは、この印影画像により、検印の確認を行う。
【００６３】
［他の実施の形態］
上述の実施の態様の他に、本発明は、次のような変形が可能である。
（１）ＯＣＲ帳票を、振込み依頼書を例に説明したが、預金引出し、新規預金等の他のＯＣＲ帳票に適用できる。更に、ＯＣＲ帳票に限らず、他の画像処理用帳票に適用できる。
（２）カラー読み取り素子を、ＣＣＤで説明したが、ＣＭＯＳセンサ等他の読み取り素子を利用できる。
（３）ドロップアウトカラーを赤色系で説明したが、他の色であっても良い。
【００６４】
以上、本発明を実施の形態により説明したが、本発明の主旨の範囲内で種々の変形が可能であり、これらを本発明の範囲から排除するものではない。
【００６５】
（付記１）帳票イメージをカラーセンサで読み取ったカラー画像データからＯＣＲ用画像データを生成する画像処理方法において、前記カラー画像データから輝度データを生成するステップと、前記カラー画像データから認識すべき文字以外の色をドロップアウトした色の画像データを抽出するステップと、前記ドロップアウトした色の画像データで前記輝度データをフィルタリングして、前記ＯＣＲ用画像データを作成するステップとを有することを特徴とする画像処理方法。
【００６６】
（付記２）付記１の画像処理方法において、前記生成ステップは、前記カラー画像データから２値化された輝度データを生成するステップであり、前記抽出ステップは、前記カラー画像データから前記ドロップアウトした色の画像データを抽出し、補完処理して、前記色のプレーンデータを作成するステップと、前記プレーンデータを２値化処理して、前記色の画像データを作成するステップとからなることを特徴とする画像処理方法。
【００６７】
（付記３）付記２の画像処理方法において、前記生成ステップは、多値の前記カラー画像データから多値の輝度データを生成するステップと、前記多値の輝度データに対し、設定された矩形フィルタ内の濃度分散値から２値化処理するステップとを有することを特徴とする画像処理方法。
【００６８】
（付記４）付記２の画像処理方法において、前記色の画像データを作成するステップは、前記プレーンデータの濃度平均に基ずいて、２値化閾値を算出するステップと、前記プレーンデータを２値化閾値で２値化するステップとを有することを特徴とする画像処理方法。
【００６９】
（付記５）付記４の画像処理方法において、前記２値化閾値を算出するステップは、前記プレーンデータのプレーンを縮小するステップと、前記縮小されたプレーンデータの濃度平均に基ずいて、２値化閾値を算出するステップとを有し、前記２値化するステップは、前記縮小されたプレーンデータを２値化閾値で２値化するステップと、前記２値化されたプレーンデータのプレーンを拡大するステップとを有することを特徴とする画像処理方法。
【００７０】
（付記６）帳票イメージを読み取り、カラー画像データを生成するイメージリーダと、前記カラー画像データから輝度データを生成し、且つ前記カラー画像データから認識すべき文字以外の色をドロップアウトした色の画像データを抽出し、前記ドロップアウトした色の画像データで前記輝度データをフィルタリングして、前記ＯＣＲ用画像データを作成する画像処理ユニットとを有することを特徴とする画像処理装置。
【００７１】
（付記７）付記６の画像処理装置において、前記処理ユニットは、前記カラー画像データから２値化された輝度データを生成し、前記カラー画像データから前記ドロップアウトした色の画像データを抽出し、補完処理して、前記色のプレーンデータを作成して、前記プレーンデータを２値化処理して、前記色の画像データを作成することを特徴とする画像処理装置。
【００７２】
（付記８）付記７の画像処理装置において、前記処理ユニットは、多値の前記カラー画像データから多値の輝度データを生成し、前記多値の輝度データに対し、設定された矩形フィルタ内の濃度分散値から２値化処理することを特徴とする画像処理装置。
【００７３】
（付記９）付記７の画像処理装置において、前記処理ユニットは、前記プレーンデータの濃度平均に基ずいて、２値化閾値を算出し、前記プレーンデータを２値化閾値で２値化することを特徴とする画像処理装置。
【００７４】
（付記１０）付記９の画像処理装置において、前記処理ユニットは、前記プレーンデータのプレーンを縮小し、前記縮小されたプレーンデータの濃度平均に基ずいて、２値化閾値を算出し、前記縮小されたプレーンデータを２値化閾値で２値化した後、前記２値化されたプレーンデータのプレーンを拡大することを特徴とする画像処理装置。
【００７５】
【発明の効果】
以上説明したように、本発明によれば、次の効果を奏する。
２次元カラーセンサに配置された各色のピクセルの各ＲＧＢの輝度であり、解像度の高い輝度データを使用するが、輝度データは、特定色をドロップアウトできないため、カラー画像データからドロップアウトカラー色の画像データを抽出し、この画像データで、高解像度の輝度データをフィルタリングすることにより、輝度データからドロップアウトカラーの部分を除去する。又、２次元カラーセンサに配置された各色のピクセルの各ＲＧＢの輝度データを、２次元カラーセンサの分解能のデータに補完処理して、これらのデータから解像度の高い輝度データを作成し、補完処理した高解像度のドロップアウトカラー色のデータでフィルタリングする。これにより、ドロップアウトカラーを除去した高解像度の輝度データをＯＣＲ画像用データとして、作成できる。これにより、高精度の文字認識が、２次元カラーセンサを用いて、実現できる。
【図面の簡単な説明】
【図１】本発明の一実施の形態の帳票処理装置のブロック図である。
【図２】図１のイメージリーダの外観図である。
【図３】図１の画像処理機構のブロック図である。
【図４】図３の画像生成部のＲプレーン生成処理の説明図である。
【図５】図３の画像生成部のＧプレーン生成処理の説明図である。
【図６】図３の画像生成部のＢプレーン生成処理の説明図である。
【図７】図３の画像生成部のＹプレーン生成処理の説明図である。
【図８】本発明の一実施の形態の読み取り処理フロー図である。
【図９】本発明の一実施の形態のＯＣＲ画像生成処理フロー図である。
【図１０】図９のＲ（２値）生成処理フロー図である。
【図１１】図１０の大津法２値化処理の説明図である。
【図１２】図９のＹ（２値）生成処理の説明図である。
【図１３】図１２のＮｉｂｌａｃｋ法の説明図である。
【図１４】図１２の文字カスレ推定処理の説明図である。
【図１５】図９のマスク合成処理のフロー図である。
【図１６】図１５のマスク合成処理の説明図である。
【図１７】図９の処理によるＯＣＲ画像生成動作の説明図である。
【図１８】図１７のＯＣＲ画像の拡大図である。
【図１９】図９の処理によるＯＣＲ画像の説明図である。
【図２０】従来技術を説明するためのＯＣＲ帳票の説明図である。
【符号の説明】
１ イメージリーダ
２ 処理装置
２０ ＣＰＵ
２１ 画像メモリ
２２ 画像処理機構
３０ 制御部
３１ 上位アプリケーション
４ 帳票
１１ カラーＣＣＤ[0001]
BACKGROUND OF THE INVENTION
The present invention relates to an image processing method and apparatus for generating OCR image data from a form image, and more particularly to an image processing method and apparatus for generating OCR image data from color image data.
[0002]
In financial institutions, post offices, etc., automatic processing of forms such as new deposits, deposit withdrawals, and transfers has become possible through character recognition (OCR) technology. For such automatic form processing, it is necessary to create OCR image data from the form image.
[0003]
[Prior art]
FIG. 20 is an explanatory diagram of an OCR form, and shows a transfer request form 4 as an example of the OCR form. The area around the entered character is surrounded by a red frame 40. The stamp 41 is in the “marking field” on the right side of the client. The seal here is the seal of the person in charge at the bank branch in system processing. The stamp 41 is also red.
[0004]
The character recognition engine of the host application recognizes the handwritten or printed characters in the entry frame 40. Therefore, there is a need for a character recognition image in which only black characters are left, in which red colors are dropped out and red colors such as an OCR form entry frame are dropped.
[0005]
The following methods are known as methods for obtaining this character recognition image.
(1) An OCR image reader that reads a form with a light source of a dropout color characteristic band is used. For example, in order to acquire form image data for character recognition, a red light source (or red filter) is irradiated, a red entry frame is dropped out, and the form is read.
(2) A color image reader equipped with a full color CCD sensor is used as an image reader for OCR character recognition. This image reader has an advantage that it can be used for various purposes such as general full-color reading and plain image reading. In this case, when reading a red dropout color conventionally used in OCR, an R plane having a low resolution is used.
[0006]
[Problems to be solved by the invention]
However, the prior art has the following problems.
(1) In the method using the image reader for OCR, the application is limited to OCR, and cannot be used for other purposes such as extraction of seals.
(2) In the method using a color image reader, since one pixel is composed of RGB, when a dropout color is used, the resolution decreases and the character recognition error rate increases. In particular, it is difficult to recognize handwritten Chinese characters.
[0007]
Accordingly, an object of the present invention is to provide an image processing method and apparatus for generating OCR image data that enables highly accurate character recognition even when a versatile color image reader is used.
[0008]
Another object of the present invention is to provide an image processing method and apparatus for generating high-resolution OCR image data from color image data.
[0009]
Still another object of the present invention is to provide an image processing method and apparatus for generating OCR image data that can be easily recognized from color image data.
[0010]
[Means for Solving the Problems]
  To achieve this object, in an image processing method for generating OCR image data from color image data obtained by reading a form image of the present invention with a color sensor, a step of generating luminance data from the color image data, and the color image Extracting image data of a color obtained by dropping out colors other than characters to be recognized from the data, filtering the luminance data with the image data of the dropped-out color, and generating the OCR image data; HaveThen, the generating step complements each of the multi-valued color image data of the color sensor covered with a filter of a color filter array of a plurality of colors in one two-dimensional sensor, and sets the resolution of the two-dimensional sensor. Creating multi-value plane data for each of the plurality of colors; generating multi-value luminance data from the generated multi-value plane data for each color; and the generated multi-value luminance data. And generating the binarized luminance data, the extracting step extracting the multi-value plane data of the dropped-out color of the resolution of the two-dimensional sensor, and extracting the multi-value of the extracted multi-value The method includes a step of binarizing plain data to create binarized image data of the color. Wherein in the binarized image data of a color that dropout by filtering the binarized luminance data, comprising the step of creating the image data for the OCR.
[0011]
  An image processing apparatus according to the present invention reads a form image and generates color image data, generates luminance data from the color image data, and drops out colors other than characters to be recognized from the color image data. An image processing unit that extracts the image data of the selected color and filters the luminance data with the image data of the dropped-out color to create the OCR image data.The image reader includes a color sensor covered with a filter of a color filter array of a plurality of colors in one two-dimensional sensor, and the image processing unit receives each of the multi-valued color image data of the color sensor. Complement processing, creating multi-value plane data of each of the plurality of colors of the resolution of the two-dimensional sensor, generating multi-value luminance data from the generated multi-value plane data of each color, Binarized luminance data is generated from the generated multivalued luminance data, and the multivalued plane data of the dropped-out color at the resolution of the two-dimensional sensor is binarized, A binarized image data is created, and the binarized luminance data is filtered by the binarized image data of the color that has been dropped out, and the OCR image data is filtered. To create the data.
[0012]
The present invention uses luminance data of color image data. The luminance data is the luminance of each RGB, and the resolution is high. However, the luminance data cannot drop out a specific color. Therefore, dropout color image data is extracted from the color image data, and the high-resolution luminance data is filtered with this image data, thereby removing the dropout color portion from the luminance data. Thereby, high-resolution luminance data from which the dropout color has been removed can be generated as OCR image data. Thereby, highly accurate character recognition becomes possible.
[0013]
In the image processing method of the present invention, the generating step is a step of generating luminance data binarized from the color image data, and the extracting step is a step of calculating the color of the dropout color from the color image data. The image data is extracted and complemented to generate the color plane data, and the plane data is binarized to generate the color image data. Data filtering processing is possible, and filtering processing is easy and fast.
[0014]
Further, in the image processing method of the present invention, the generation step includes a step of generating multi-value luminance data from the multi-value color image data, and a step in a rectangular filter set for the multi-value luminance data. By including the step of binarization processing from the density dispersion value, it is possible to perform binarization optimally locally, and to prevent missing characters that are information portions.
[0015]
Furthermore, in the image processing method of the present invention, the step of creating the color image data includes: calculating a binarization threshold value based on an average density of the plane data; and converting the plane data into the binarization threshold value. And the step of binarizing can generate mask data in which the character portion of the luminance data is not missing.
[0016]
Furthermore, in the image processing method of the present invention, the step of calculating the binarization threshold includes the step of reducing the plane of the plane data and the binarization threshold based on the density average of the reduced plane data. And the step of binarizing includes the step of binarizing the reduced plane data with a binarization threshold and the step of enlarging the plane of the binarized plane data The mask data can be generated at high speed.
[0017]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, the present invention will be described by dividing it into a form recognition device, an OCR image generation process, and other embodiments.
[0018]
[Form recognition device]
1 is a block diagram of a form recognition apparatus according to an embodiment of the present invention, FIG. 2 is an external view of a color image reader of FIG. 1, FIG. 3 is a block diagram of an image processing mechanism of FIG. FIG. 7 is an explanatory diagram of the generation processing of each plane of FIG. 3, and FIG. 8 is an image processing flowchart of FIG.
[0019]
As shown in FIG. 1, the form processing apparatus includes a color image reader 1 and a processing apparatus 2. As shown in FIG. 2, the color image reader 1 includes a bet (form pedestal) 10 for setting a form 4 to be read and an overhead 15 for reading an image of the form of the bet 10. As shown in FIG. 1, the overhead 15 is provided with a two-dimensional color CCD (Charge Coupled Device) 11. The reading circuit 12 controls the reading operation of the color CCD 11. The amplifier circuit 13 amplifies the read output of the color CCD 11. The A / D (analog / digital) converter 14 converts the amplified analog read signal into a digital signal. The image reader 1 can perform color reading and is non-contact. Also, since a two-dimensional sensor is used, the reading speed is fast. In FIG. 2, a serial number printer 16 is provided as an option.
[0020]
The processing device 2 is composed of a desktop personal computer and is connected to a host computer. The processing device 2 includes a data processing unit (CPU) 20, an image memory 22, an image processing mechanism 21, a display device, and a file. The data processing unit 20 includes a CPU, a memory, and the like, and executes a form image process. Software executed by the CPU 20 is indicated by blocks 30 to 32 in the CPU 20. The host application 31 is a form processing program. The control unit 30 controls the image reader 1 and the image processing mechanism 21 in order to perform image processing requested by the host application 31. The OCR engine 32 is activated by the upper application 31 and performs character recognition processing.
[0021]
The image processing mechanism 21 will be described with reference to FIGS. As shown in FIG. 3, RGB data is output from the color CCD 11, and the RGB data of one screen (form) is held in the image memory 22. The image generation unit 50 generates a Y plane image, an R plane image, a G plane image, and a B plane image from the RGB data in the image memory 22, respectively.
[0022]
As shown in FIG. 4, the color CCD 11 is formed by covering each normal CCD with an RGB filter (on-chip filter), and has a checkered structure. One pixel is composed of four elements R, G (Gr, Gb), and B. An image acquired via the RGB filter is output from the CCD 11 as an RGB image. The resolution is 240 dpi.
[0023]
The red filter R has a peak wavelength of 650 nm, the green filter G has a peak wavelength of 540 nm, and the blue filter B has a peak wavelength of 470 nm.
[0024]
The image generation unit 50 generates an image (multivalue) of R, G, B planes from the output RGB of the color CCD 11 from the checkered CCD array structure, and generates an image Y of the Y plane.
[0025]
The R image passes through the red filter R, and has a characteristic of dropping out the red system. For example, in an OCR form in which an entry frame is preprinted in a red color, the red entry frame is removed, and only a color other than red, such as a black entry character, is generated as an R image.
[0026]
The G image passes through the green filter G and has a feature of dropping out a green color. The ability to drop out is low. For example, since the imprint vermilion is stamped in red, the G image is often used as a reliable method for reading the imprint.
[0027]
The B image passes through the blue filter B and has a feature of dropping out a blue color. Rather than using it alone, it is used as the B plane of the color image RGB.
[0028]
As described above, since one pixel is composed of four elements, in order to obtain a plane image having the same resolution 240 dpi as the input original image 240 dpi, the plane is divided and pixel interpolation processing is performed. FIG. 4 shows the complement processing of the R plane, and the eight pixels R1 to R8 around the pixel R are created by the following equations.
[0029]

Next, FIG. 5 shows the G plane complement processing, and the four pixels G1 to G4 around the pixel G are created by the following equations.
[0030]

Further, FIG. 6 shows B plane complementation processing, and the eight pixels B1 to B8 around the pixel B are created by the following equations.
[0031]

Next, generation of a Y plane image (240 dpi) will be described with reference to FIG. The Y image is called a monochrome image. Since each RGB image is an image read through a unique filter, a specific color component is dropped out, whereas a Y image is luminance data that directly outputs each color component. This Y plane image can be generated from an RGB plane image. That is, Y is generated from RGB by the following ratio calculation formula.
[0032]
Y = 0.299 × R + 0.587 × G + 0.114 × B
As is apparent from the above formula, the reason why the ratio of the G plane image is high is that many color components are included in the G plane. In this way, a monochrome image can be generated by individually separating and combining RGB.
[0033]
Returning to FIG. 3, the binarization processing unit 51 binarizes the multivalued Y, R, and G images to generate a Y, R, and G image (binarization). The aforementioned Y, R, and G plane images are composed of multi-value image data having 256 gradations. When performing recognition and collation processing by a computer such as character recognition and seal collation, it is necessary to consider performance and the like, and the binarization processing unit 51 converts 256-level multi-valued image data of “0 or 1” to 2 Convert to valued image data.
[0034]
As described later, the OCR image generation unit 52 generates OCR image data in the following procedure. First, since each plane image is filtered by each light receiving sensor, the R plane is used if it is a red dropout color. On the R plane, the dropout color portion is dropped out and the character portion has a low resolution (equivalent to 120 dpi). The R plane image is reduced to ¼ size by OR thinning-out so that there is no missing information in the character part. Dynamic two-tone separation (Otsu method binarization processing) is performed on the background portion and the character portion in the reduced plain image. The Otsu method binarized plain image is returned to the size before reduction. A high-definition image can be acquired with a filter with a later Y signal (luminance signal) by squashing the character portion.
[0035]
A Y signal (luminance signal) is used to make the character portion high definition. Since the Y signal is an arithmetic synthesis of all plane images, it is originally high definition (equivalent to 240 dpi). Floating binarization processing (Niblack method binarization) is performed on the Y signal, and binarization is performed so as to output information (character blur) as much as possible. In this case, the dropout color portion becomes a black pixel (not dropped out) by the floating binarization process, but unnecessary information can be erased later by performing filter synthesis with the plain image. A high-definition dropout color image is generated by filtering the R plane binary and the Y signal (luminance signal) binary. In this case, an image can be generated by a simple AND mask operation.
[0036]
As for the binarization processing of the Otsu method, the literature, Nobuyuki Otsu: “Automatic threshold selection method based on discrimination and least squares criterion”, IEICE Transactions, J63-D, vol.4, pp .349-356 (1980). The Niblack method binarization is described in detail in the literature, W. Niblack: “An Introduction to Digital Image Processing”, pp. 115-116, Englewood Cliffs, NJ: Prentice Hall (1986). .
[0037]
The form image processing will be described with reference to FIG.
(S100) The host application 31 instructs the control unit 30 to read. The reading instruction includes a reading instruction and a request image. The request image is specified as a character recognition image binary or a seal image binary. Of course, this reading instruction is after the form is mounted on the bed 10 of the image reader 1.
(S110) The control unit 30 instructs the reading circuit 12 of the image reader 1 to read the color CCD 11. As a result, the color CCD 11 starts reading and outputs RGB data via the amplifier 13 and the A / D converter 14.
(S120) This RGB data is stored in the image memory 22. Thereby, the reading is completed, and the operator can remove the form.
(S130) After this storage, according to an instruction from the control unit 30, the image processing unit 21 performs the above-described image processing for generating an image requested by the host. Thereby, each plane image is stored in a memory (not shown) of the image processing unit 21.
(S140) The control unit 30 notifies the upper application 31 of the upper request image from the plane memory. The OCR engine 32 performs character recognition processing on the OCR image.
[0038]
In this manner, it is possible to generate dropout color OCR image data using high-resolution luminance data. Due to the dropout color, the character frame and the like are removed, and a high-resolution character image is obtained. Therefore, character recognition becomes accurate, and complex character recognition such as handwritten Chinese characters can be made accurate.
[OCR image generation]
Hereinafter, an example of creating a 240 dpi red dropout color image will be described as an example.
[0039]
9 is a flowchart of OCR image processing, FIG. 10 is an explanatory diagram of R (binary) data generation operation, FIG. 11 is an explanatory diagram of Otsu method binarization processing, and FIG. 12 is a binary of a luminance signal. FIG. 13 is an explanatory diagram of the Niblack binarization process, FIG. 14 is an explanatory diagram of an estimation process of the blurred portion of the character, FIG. 15 is an explanatory diagram of the filter process, and FIG. It is operation | movement explanatory drawing of a filter process.
[0040]
As shown in FIG. 9, even if pixel complement processing (division for each plane and creation of luminance signal) is performed by inputting the original image, the R plane 59 has a small amount of information, so the number of pixels is 240 dpi, but the resolution is high. Has a low resolution (equivalent to 120 dpi).
[0041]
The OR operation thinning 60 is performed on the R plane image 59 corresponding to ¼. The purpose of OR thinning is to prevent the luminance signal information (characters) from being lost and to binarize the R plane 59 at high speed when filter masking the subsequent luminance signal image. As shown in FIG. 10, the pixel Ra of the ¼ plane selects the second highest density pixel in the rectangular matrix of the R plane pixels R1 to R4.
[0042]
Next, the Otsu method binarization processing 61 is performed on the R-plane image reduced to ¼. Since the R-plane image is dropped out, binarization is performed using a dynamic slice (threshold) between the background portion and the character portion. As shown in FIG. 11, the Otsu method binarization calculates a histogram (number of pixels of each density value) in an image, and calculates the occurrence probability distribution and the overall density average from the histogram (density distribution) by the following formula. Ask. The binarization threshold is calculated by obtaining the inter-class variance from the occurrence probability distribution.

[0043]
The density value having the maximum inter-class variance value is determined as a binarized slice level, and the density value of each pixel is sliced at this slice level and binarized.
Further, the binarized R-plane image is returned to the original size. That is, the image data 63 of the original size 240 dpi is returned to the 240 dpi luminance signal and the filter mask by performing the 4 × enlargement processing 62.
[0044]
On the other hand, as shown in FIGS. 9 and 12, Niblack method binarization 65 is performed on the luminance signal Y multilevel 64. Since the luminance signal is a 240 dpi high-definition image, all the information is output by performing binarization in a floating manner. However, since the Niblack method binarization is a process that takes time to process on software, only the blurred portion of the character is estimated (FIG. 14), and the Niblack method binarization is performed only on that portion.
[0045]
As shown in FIG. 13, in the Niblack method binarization, a rectangular filter (the size of the filter is arbitrary) is set, and the background portion or the information portion is discriminated from the density dispersion value g2 in the filter. If the density dispersion value g2 is equal to or greater than the dispersion threshold (σmin), the density threshold S of the pixel of interest [k] is obtained, and the density value of the pixel of interest is compared with the threshold S to determine whether the pixel is white or black. judge. Since the threshold is set for the local region, there is little loss of information.
[0046]
In this embodiment, since the handwritten character recognition image for OCR is generated, a 7 × 7 filter is set (the stroke width of the handwritten character is about 3 dots, so a 7 × 7 filter is desirable).
[0047]
In addition, since a rectangular area is set for each pixel of interest (k) for which a threshold value is to be calculated, three pixels in the upper, lower, left, and right directions, if the pixel of interest is a background (in the filter and only the background), the background threshold σmin in FIG. The background is determined by the determination formula (1). In addition, since the absolute value of density is not observed, even if the character stroke is completely covered in the 7 × 7 filter, it is regarded as the background, but the character stroke portion is determined by “estimation of character blur” described later. The inside can be complemented with black. When the character stroke and the background are included in the filter, the density dispersion value g2 is equal to or greater than the background threshold σmin, and the threshold S of the pixel of interest k in FIG. 13 is obtained from Expression (2) in FIG. If the pixel of interest has a density equal to or lower than the obtained threshold value S, the pixel is a black pixel.
[0048]
Further, the root (g) of the formula (2) in FIG.²) Is multiplied by a coefficient (-0.1). It is also possible to suppress noise around the character by setting the threshold value S as an undertone.
[0049]
Next, the Otsu method binarization threshold value calculation method used in the R-plane binarization in order to estimate the blurred portion of the character of the luminance signal is summarized. As shown in FIG. 14, a dynamic two-tone separation threshold value Ot is obtained from the entire luminance signal image by the Otsu method, and a low density direction (dark density) and a high density direction (dark density) are centered on the threshold value Ot. ) For the ratios a and b. That is, the character blur portion estimation area is calculated according to the following equation.
[0050]
Scratch range coefficient: a = 0.10 to 0.90 b = 0.10 to 0.90
Average density in black area: b # ave = Average density of density distribution 0 to Ot
White area density average: w # ave = density distribution Ot ~ 255 density average
Lower limit of blurred range bt = ((1.0-a) x ot) + (a x b # ave)
Upper limit of blurred range wt = ((1.0-b) x ot) + (b x w # ave)
Niblack binarization is performed on the pixel of interest with the density value in the above estimated area (wt to bt), and other pixels in the b # ave area are determined to be black and pixels in the w # ave area are determined to be white. Therefore, it is possible to perform a large speed processing. If the ratio of a and b is increased, the portion of local binarization increases. However, if the ratio of a and b is increased too much, it becomes useless processing when binarizing a handwritten character. In this example, the coefficient of the blur range is set to a = 0.35, b = 0.62 (5-8% of the whole is Niblack processing: the ratio varies depending on the number of characters entered), and the image subjected to 100% processing The same level of results were obtained.
[0051]
Then, as shown in FIGS. 9, 15, and 16, mask filter synthesis processing 67 is performed on the luminance signal binary data 66 and the R plane binary data 63. That is, by combining the binarized images with a simple AND mask, the dropout portion is the R plane image 63 and the character portion (information portion) is the luminance signal as shown in the composite image 68 of FIG. Information can be extracted from the image 66.
[0052]
In this way, luminance data having a high resolution and a luminance of each RGB is used. However, the luminance data cannot drop out a specific color. Therefore, dropout color image data is extracted from the color image data, and the high-resolution luminance data is filtered with this image data, thereby removing the dropout color portion from the luminance data. Thereby, high-resolution luminance data from which the dropout color has been removed can be generated as OCR image data. Thereby, highly accurate character recognition becomes possible.
[0053]
In addition, binarized luminance data is generated from the color image data, and the image data of the color dropped out from the color image data is extracted and complemented to create color plane data. By creating the color image data by performing the binarization process, it becomes possible to perform the filtering process of the binary data, and the filtering process becomes easy and fast.
[0054]
Further, multi-value luminance data is generated from the multi-value color image data, and the multi-value luminance data is locally processed by the Niblack method in which binarization processing is performed from the density dispersion value in the set rectangular filter. Optimal binarization can be performed, and missing characters that are information portions can be prevented.
[0055]
Further, the binarization threshold value is calculated based on the average density of the plane data, and the binarization process such as the Otsu method for binarizing the plane data with the binarization threshold value eliminates the loss of the character portion of the luminance data. No mask data can be generated.
[0056]
Further, the step of calculating the binarization threshold includes the step of reducing the plane of the plane data and the step of calculating the binarization threshold based on the density average of the reduced plane data. The step includes the step of binarizing the reduced plane data with a binarization threshold and the step of enlarging the plane of the binarized plane data, so that mask data can be generated at high speed.
[0057]
As an example, the image data of the transfer request form 4 as an example of the aforementioned OCR form in FIG. 20 will be described with reference to FIGS. 17 and 18. In this OCR form 4, the entered characters are surrounded by a red frame 40. The stamp 41 is in the “marking field” on the right side of the client. The seal here is the seal of the person in charge at the bank branch in system processing. The stamp 41 is also red.
[0058]
As shown in FIG. 17, in the Y signal binary data 66 of the OCR form 4 described above, an entry frame 40, entry characters, and a seal 41 appear. On the other hand, in the R plane binary image data 63, the red entry frame 40 and the stamp 41 are dropped out, and the entry characters appear. The Y signal binary data 66 is subjected to mask composition processing 67 with the R plane binary data 63 to obtain OCR image data 68. An enlarged view of the OCR image data 68 is shown in FIG. That is, it can be seen that the dropout color portion is removed from the high-resolution Y signal.
[0059]
The characteristics of the OCR image data will be described with reference to FIG. FIG. 19 shows a Y signal binary image 66, an R plane binary image 63, and an OCR image 68 of a handwritten character of “Hokkaido Takushoku” surrounded by an entry frame. The Y signal binary image 66 has a high resolution, but an entry frame remains. In the R plane image 63, the entry frame is removed, but the characters have a low resolution. The OCR image 68 has a high resolution, and the entry frame is also removed.
[0060]
In the system of FIG. 1 described above, the upper application 31 designates the sequential reading mode as a reading instruction, and designates the character recognition image binary value for the first time and the imprint image binary value for the second time for the request image.
[0061]
In response to the request “character recognition image binary”, the control unit 30 transfers the OCR image (binary) binarized by the OCR processing unit 52 to the upper level. Since this OCR image (binary) drops out red, the red system such as an OCR form entry frame is dropped and only black characters remain. For this reason, it is an optimal image as a character recognition image. As described above, the OCR image (binary) represents a high-resolution character image, the character recognition entry frame 40 is all dropped out, and the stamp 41 is also dropped out.
[0062]
Subsequently, the G image (binary) binarized by the binarization processing unit 51 is transferred to the upper level in response to the image request “binary impression image binary”. This G image (binary) has a feature that expresses the red color of the seal imprint as the image with the highest sensitivity. For this reason, it is an optimal image as an imprint image. In the G image (binary), the imprint 41 appears clearly as an image. Of course, at this time, since the red color image is output, the character recognition entry frame 40 and the like remain in the image. The upper 31 cuts out the seal imprint area from the G image and transfers it to the host. The host confirms the seal using this imprint image.
[0063]
[Other embodiments]
In addition to the above-described embodiments, the present invention can be modified as follows.
(1) Although the OCR form has been described by taking the transfer request form as an example, it can be applied to other OCR forms such as deposit withdrawal and new deposit. Furthermore, the present invention can be applied not only to OCR forms but also to other image processing forms.
(2) Although the color reading element has been described as a CCD, other reading elements such as a CMOS sensor can be used.
(3) Although the dropout color has been described as red, other colors may be used.
[0064]
As mentioned above, although this invention was demonstrated by embodiment, a various deformation | transformation is possible within the range of the main point of this invention, and these are not excluded from the scope of the present invention.
[0065]
(Appendix 1) In an image processing method for generating image data for OCR from color image data obtained by reading a form image with a color sensor, a step of generating luminance data from the color image data, and a character to be recognized from the color image data A step of extracting image data of a color in which a color other than the above is dropped out, and a step of creating the OCR image data by filtering the luminance data with the image data of the color that has been dropped out. Image processing method.
[0066]
(Supplementary Note 2) In the image processing method according to Supplementary Note 1, the generation step is a step of generating binarized luminance data from the color image data, and the extraction step is the dropout from the color image data. Extracting color image data, performing complementary processing to create plane data of the color, and binarizing the plane data to generate image data of the color An image processing method.
[0067]
(Supplementary note 3) In the image processing method according to supplementary note 2, the generation step includes a step of generating multi-value luminance data from the multi-value color image data, and a rectangular filter set for the multi-value luminance data. And a binarization process from the density dispersion value in the image processing method.
[0068]
(Supplementary Note 4) In the image processing method according to Supplementary Note 2, the step of creating the color image data includes a step of calculating a binarization threshold based on a density average of the plane data, and binarizing the plane data. An image processing method comprising: binarizing with a threshold value.
[0069]
(Supplementary Note 5) In the image processing method of Supplementary Note 4, the step of calculating the binarization threshold value is based on a step of reducing the plane of the plane data and a density average of the reduced plane data. A binarization step, wherein the binarization step binarizes the reduced plane data with a binarization threshold and enlarges the plane of the binarized plane data. And an image processing method.
[0070]
(Supplementary Note 6) An image reader that reads a form image and generates color image data, an image of color that generates luminance data from the color image data, and drops out colors other than characters to be recognized from the color image data An image processing apparatus comprising: an image processing unit that extracts data and filters the luminance data with the image data of the color that has been dropped out to generate the OCR image data.
[0071]
(Supplementary note 7) In the image processing apparatus according to supplementary note 6, the processing unit generates binarized luminance data from the color image data, extracts image data of the dropped-out color from the color image data, An image processing apparatus, characterized in that the color plane data is generated by complementing, and the plane data is binarized to generate the color image data.
[0072]
(Supplementary note 8) In the image processing apparatus according to supplementary note 7, the processing unit generates multi-valued luminance data from the multi-valued color image data, and the multi-valued luminance data is stored in a set rectangular filter. An image processing apparatus that performs binarization processing from a density dispersion value.
[0073]
(Supplementary note 9) In the image processing apparatus according to supplementary note 7, the processing unit calculates a binarization threshold based on a density average of the plane data, and binarizes the plane data with a binarization threshold. An image processing apparatus.
[0074]
(Supplementary note 10) In the image processing apparatus according to supplementary note 9, the processing unit reduces the plane of the plane data, calculates a binarization threshold based on a density average of the reduced plane data, and reduces the reduction An image processing apparatus characterized in that the binarized plane data is binarized with a binarization threshold, and then the plane of the binarized plane data is expanded.
[0075]
【The invention's effect】
  As described above, the present invention has the following effects.
  Of each color pixel placed in the 2D color sensorThe luminance data of each RGB is high-resolution luminance data. However, since the luminance data cannot drop out a specific color, the image data of the dropout color is extracted from the color image data. By filtering the resolution luminance data, the dropout color portion is removed from the luminance data.Also, the RGB luminance data of each color pixel arranged in the two-dimensional color sensor is complemented to the resolution data of the two-dimensional color sensor, and high-resolution luminance data is created from these data, and the complementary processing is performed. Filter with high resolution dropout color data.As a result, high-resolution luminance data from which the dropout color is removed can be created as OCR image data. This enables high-precision character recognition.Realized using a two-dimensional color sensorit can.
[Brief description of the drawings]
FIG. 1 is a block diagram of a form processing apparatus according to an embodiment of the present invention.
2 is an external view of the image reader of FIG. 1. FIG.
3 is a block diagram of the image processing mechanism of FIG. 1. FIG.
4 is an explanatory diagram of R plane generation processing of the image generation unit in FIG. 3; FIG.
FIG. 5 is an explanatory diagram of G plane generation processing of the image generation unit in FIG. 3;
6 is an explanatory diagram of a B plane generation process of the image generation unit of FIG. 3;
7 is an explanatory diagram of a Y plane generation process of the image generation unit of FIG. 3;
FIG. 8 is a flowchart of a reading process according to the embodiment of this invention.
FIG. 9 is an OCR image generation processing flowchart according to an embodiment of the present invention.
10 is an R (binary) generation process flow diagram of FIG. 9; FIG.
FIG. 11 is an explanatory diagram of the Otsu method binarization processing of FIG. 10;
12 is an explanatory diagram of the Y (binary) generation process of FIG. 9. FIG.
13 is an explanatory diagram of the Niblack method of FIG. 12. FIG.
FIG. 14 is an explanatory diagram of the character blur estimation process in FIG. 12;
FIG. 15 is a flowchart of the mask composition process of FIG. 9;
16 is an explanatory diagram of the mask composition process of FIG. 15;
17 is an explanatory diagram of an OCR image generation operation by the process of FIG.
18 is an enlarged view of the OCR image of FIG.
FIG. 19 is an explanatory diagram of an OCR image obtained by the process of FIG.
FIG. 20 is an explanatory diagram of an OCR form for explaining the prior art.
[Explanation of symbols]
1 Image reader
2 processing equipment
20 CPU
21 Image memory
22 Image processing mechanism
30 Control unit
31 Top applications
4 forms
11 Color CCD

Claims (2)

In an image processing method for generating image data for OCR from color image data obtained by reading a form image with a color sensor,
Generating luminance data from the color image data;
Extracting image data of a color obtained by dropping out a color other than the character color to be recognized from the color image data;
Filtering the luminance data with the image data of the dropped-out color to create the OCR image data,
The generating step includes
Each of the plurality of color image data of the resolution of the two-dimensional sensor is complemented by each of the multi-valued color image data of the color sensor covered with a filter of a color filter array of a plurality of colors on one two-dimensional sensor. Creating multi-value plane data;
Generating multi-valued luminance data from the generated multi-value plane data of each color;
Generating binarized luminance data from the generated multi-valued luminance data;
The extraction step includes
Extracting the drop-out color multi-value plane data of the resolution of the two-dimensional sensor;
The extracted multi-value plane data is binarized to extract binarized image data of the color;
The creating step includes
Filtering the luminance data in which the binarized with binarized image data of a color that is the drop-out, Ri Do the step of creating image data for the OCR,
The step of creating the binarized image data of the color includes:
Reducing the plane data;
Calculating a binarization threshold based on a density average of the reduced plane data;
Binarizing the reduced plane data with the binarization threshold;
And enlarging the plane of the binarized plane data . An image reader that reads form images and generates color image data;
Luminance data is generated from the color image data, and image data of a color in which a color other than a character color to be recognized is dropped out of the color image data is extracted. The luminance data is extracted from the image data of the dropped out color. An image processing unit for filtering and creating the OCR image data,
The image reader has a color sensor covered with a filter of a plurality of color filter arrays in one two-dimensional sensor,
The image processing unit includes:
The multi-value color image data of the color sensor is complemented to generate multi-value plane data of each of the plurality of colors having the resolution of the two-dimensional sensor, and the generated multi-value plane of each color Multi-value luminance data is generated from the data, binary luminance data is generated from the generated multi-value luminance data, and the multi-value of the dropped-out color of the resolution of the two-dimensional sensor is generated. The binarization processing of the plane data is performed to create the binarized image data of the color, and the binarized luminance data is filtered by the binarized image data of the dropout color. , Create the OCR image data ,
Further, in order to create binary image data of the color, the plane data is reduced, a binarization threshold value is calculated based on a density average of the reduced plane data, and the reduced plane An image processing apparatus characterized in that data is binarized with the binarization threshold and the plane of the binarized plane data is enlarged .

Images