JP4733859B2 - Location information recognition device and sorting device - Google Patents

Description

総合評価では、複合類似度の認識結果と構造特徴マッチングの認識結果の全組み合わせについてそれぞれ評価値を算出し、その上位２候補を、それまでの前文字列に対する答えとする。具体的には、（例１）の場合、前文字列が「１」「２」の場合についてそれぞれ
（複合類似度の認識結果、構造特徴マッチングの認識結果）
＝（０−１，０−１）、（０−１，８−１）、（６−２，０−１）、
（６−２，８−１）、（０−１，×）、（６−１，×）、（×，０−１）、
（×，８−１）
の全組み合わせ（×は、候補がない状態）について、以下で示す第２２〜第２５のステップの操作を行ってそれぞれの総合評価値を算出し、評価値の高い方から１候補または２候補（候補数の選択の詳細は、第２６のステップを参照）を、その前文字列候補に対する答えとして採用する。
【００３６】
尚、第２２〜第２５のステップでは、前文字列が「１」で、（複合類似度の認識結果、構造特徴マッチングの認識結果）＝（０−１，０−１）の場合を例に挙げて説明する。
【００３７】
第２２のステップでは、文字の出現率を求める。図５を用いて説明する。出現率評価テーブル４１には、各桁での文字の出現率が、前文字列候補に依存する形で設定されている。図５は、１桁目と２桁目の出現率評価テーブル４１の例である。この場合、２桁目に「０」が出る確率は、１桁目の文字が「１」のときは６７％であるが、１桁目の文字が「２」の時は１００％となることを示す。同様に、１桁目の文字が「１」のとき、２桁目で「１」が出る確率は１８％、「２」が出る確率は１５％、その他は０％であることが求められる。自動調整処理では、ある桁の出現率を求める場合、前文字列と現在評価する桁の構造特徴マッチングの認識結果とを用いて出現率評価表としての出現率評価テーブル４１を探索する。
【００３８】
（例１）の場合、１桁目候補が「１」で、２桁目の構造特徴マッチング候補は「０」「８」である。よって、「０」の出現率６７％が得られる。
次に、処理する桁をＮ桁目（１≦Ｎ≦７）、構造特徴マッチングの認識結果をＣＮ、前文字列Ｃ１、Ｃ２…ＣＮ−１＝Ｃｎとした場合の出現率の求め方について、図１０に示すフローチャートを参照しつつ説明する。
【００３９】
すなわち、Ｎ桁目の出現率を求める場合、ｎを１に設定し（ＳＴ３１）、ｎがＮと一致するかチェックし（ＳＴ３２）、一致する場合、１桁目の出現率評価テーブル４１から構造特徴マッチングの認識結果ＣＮの出現率（Ａ１）を得る（ＳＴ３５）。
【００４０】
上記ステップ３２のチェックの結果、一致しない場合、１桁目の出現率評価値により次桁へのポインタを求め（ＳＴ３３）、ｎに１を加算し（ｎ＋＋）（ＳＴ３４）、ステップ３２に戻る。
このステップ３２のチェックの結果、一致する場合、上記ステップ３５に進み、Ｎ桁目の出現率評価テーブル４１から構造特徴マッチングの認識結果ＣＮの出現率（Ａ１）を得、一致しない場合、上記ステップ３３に進む。
【００４１】
第２３のステップでは、図１１のフローチャートに示すように、調整パターン番号を求める。
結合テーブル４２から、構造特徴マッチングの認識結果、複合類似度の認識結果が一致する行を探し（ＳＴ４１、４２、５０）、更に、「相手の条件」および「自分の条件」を満たす（ＳＴ４３〜４７）最初の行（Ｉ）のパターン番号を調整パターン番号（Ａ２）として出力する（ＳＴ４８）。
【００４２】
最終行まで達した際、最終行のパターン番号を調整パターン番号（Ａ２）として出力する（ＳＴ４９）。
「相手文字」とは、誤読しやすいと想定される相手の数字のことであり、これはシミュレーションや過去の調整経験から得られる既知の情報から設定されている。たとえば、構造特徴マッチングの認識結果で０−１の結果が得られたパターンは「６」に誤読することが多い、等といった情報が、結合テーブル４２に盛り込まれている。
【００４３】
例の場合、構造特徴マッチングの認識結果が０−１、複合類似度の認識結果が０−１であり、この時点では、図６の１行目から４行目が該当し、１行目から順に条件が一致するかを検索する。
１行目：
出現率評価表としての出現率評価テーブル４１より、現在の桁では「６」の出現率が０％であり、これは「相手の条件」に一致しない。
【００４４】
２行目：
同様に、「８」の出現率は０％で、「相手の条件」が一致しない。
３行目：
相手の条件が（０＜６）で一致し、自分の条件が（（６７−０）＞２０）で一致し、この行が選ばれる。以上の結果、この例の場合、調整パターン番号３が選ばれる。
【００４５】
第２４のステップでは、頻度レベルを選択する。
構造特徴マッチングの認識結果の出現率によって、強強化、強化、標準、弱化、弱弱化の５段階の頻度レベルを定める。レベル選択テーブル４３では、構造特徴マッチングの認識結果のエントリ毎に、各レベルの間の閾値を設定している。構造特徴マッチング法のエントリが文字の特性を反映しているので、エントリの信頼性の高低によって頻度レベルの出方をかえることが可能となる。
【００４６】
今、図７に示すレベル選択テーブル４３が与えられたとき、（例１）の場合では、第２３のステップで構造特徴マッチングの認識結果が０−１、出現率が６７％（＞２５％）が得られており、この頻度レベルは強強化となる。
【００４７】
次に、頻度レベルの選択について、図１２に示すフローチャートを参照しつつ説明する。
すなわち、構造特徴マッチングの認識結果としての文字とエントリと一致するものをレベル選択テーブル４３の１行目から順次チェックし（ＳＴ５１、５２、５４、５５）、一致した場合に、出現率が強強化と強化の頻度レベルを区別する閾値よりも大きい場合に（ＳＴ５６）頻度レベル（Ａ３）を強強化と判別し（ＳＴ５７）、出現率が強化と標準の頻度レベルを区別する閾値よりも大きい場合に（ＳＴ５８）頻度レベル（Ａ３）を強化と判別し（ＳＴ５９）、出現率が標準と弱化を区別する閾値よりも大きい場合に（ＳＴ６０）頻度レベル（Ａ３）を標準と判別し（ＳＴ６１）、出現率が弱化と弱弱化を区別する閾値よりも大きい場合に（ＳＴ６２）頻度レベル（Ａ３）を弱化と判別し（ＳＴ６３）、出現率が弱化と弱弱化を区別する閾値よりも小さい場合に（ＳＴ６２）頻度レベル（Ａ３）を弱弱化と判別する（ＳＴ６４）。
【００４８】
また、ステップ５２により最後の行までチェックしても不一致の際、例外処理として頻度レベル（Ａ３）を標準とする（ＳＴ５３）
第２５のステップでは、総合評価値を算出する。
【００４９】
第２３、第２４のステップで求められた調整パターン番号と頻度レベル、および、複合類似度の信頼度を用いて、レベルテーブル４４から評価値を選択する。
図８に示すレベルテーブル４４を用いた場合、（例１）の場合では、複合類似度の認識結果が０−１（３）、パターン番号３、頻度レベルが強強化であるから、総合評価値は「６」となる。
【００５０】
次に、総合評価値の算出について、図１３に示すフローチャートを参照しつつ説明する。
すなわち、第２３のステップで求められた調整パターン番号（Ａ２）と一致するものをレベルテーブル４４の１行目から順次チェックし（ＳＴ７１、７２、７３、７６）、一致した場合に、第２４のステップで求められた頻度レベル（Ａ３）と複合類似度の信頼度から総合評価値（Ａ４）を選択し、出力する（ＳＴ７４）。
【００５１】
また、ステップ７２により最後の行（Ｉ）までチェックしても不一致の際、例外処理として総合評価値「０」（Ａ４＝０）を出力する（ＳＴ７５）。
【００５２】
第２６のステップでは全ての前文字列に対して総合評価処理を行う。同じ複合類似度の認識結果と構造特徴マッチングの認識結果が入力されても、前文字列候補によって総合評価値は変化する可能性がある。
（例１）の場合、前文字列「１」について全ての複合類似度の認識結果、構造特徴マッチングの認識結果の組み合わせで評価を行った後、前文字列「２」の場合について、同様に総合評価を行う（ＳＴ２７からＳＴ２９、ＳＴ２２からＳＴ２６）。
【００５３】
結果的に、７桁目まで上記総合評価を行うと（ＳＴ２７）、複数の郵便番号候補が得られる。各桁で求められた総合評価値を候補ごとにそれぞれ加算し、それを郵便番号候補の総合評価点とする（ＳＴ１３）。この総合評価点の高いものを上位２候補選択し、最終結果とする（ＳＴ１４）。
【００５４】
以上のような総合評価算出操作を単純に行い、存在する前文字列に対して各桁で総合評価点の高いほうから第２位までを選んで処理すると、例えば３桁目までで、図１４に示すような２分木の形となる（括弧内の数字は総合評価点）。この結果、４桁目の総合評価では、８種の前文字列候補が存在する。
【００５５】
以上の操作を７桁まで行った場合、得られる候補は１２８通りであり、各桁ごとでの総合評価を行う回数の総和は１２７回、しかも、その１２７回のそれぞれで複合類似度の認識結果と構造特徴マッチングの認識結果の全ての組み合わせを評価することになり、計算量が膨大なものとなる。そこで、以下のようなルールで計算量を削減することもできる。
【００５６】
１．１桁目は２候補までを選択。
２．２桁目以降では、図１４に示した２分木のうち、総合評価点の高い方のノードに接続するものだけ２候補選択し、他ノードでは１候補のみ選択する。
この操作で第３桁目まで総合評価を行った場合、図１５のようになり、４桁目では、前文字列候補４個に対して総合評価を行う。この方法で７桁目まで処理した場合、最終的な候補は８候補まで絞られる。この８候補の中で、各桁毎の評価点の総和が最も高いものを順に２位まで選択し、最終的な郵便番号候補とする。
【００５７】
なお、上記実施例では、文字認識の手法について複合類似度法と構造特徴マッチング法の２つの認識手法に適用した場合について説明したが、本発明はこれに限定されるものでなく、たとえば、一方もしくは両方の認識手法を別の手法に置き換えた場合にも同様に適用できる。また、認識手法が１種類である場合にも同様に適用できる。
【００５８】
さらに、使用するテーブル類もこれに限定されるものでない。上記実施例では、経験やシミュレーションなどによって得られた既知の誤読情報を、テーブルに反映させるための方法について説明したが、より単純に頻度を反映させるためには、たとえば、認識結果から得られた類似度に出現率をかけたものを総合評価値とする、などといった方法も考えられる。
【００５９】
次に、出現率の求め方について説明する。
出現率評価表としての出現率評価テーブル４１は、各桁での０〜９の出現頻度が、前文序列に依存した形で設定されたファイルであり、図１６に示すような、稼働情報ファイル１１ａから作成する。この上記制御部１１内に存在し、実際に区分機１で処理した（認識した）郵便番号とその通数を集計して保存したものである。
【００６０】
上記稼働情報ファイル１１ａは郵便局毎の複数台の区分機を管理する機器に存在し、実際に各区分機で処理した（認識した）郵便番号とその通数を集計して保存したものであっても良い。
従来は作業者が経験的に捉えていた文字の出現率が、実際の機械から得られた統計的な情報に置き換えられ、より正確に文字認識処理に反映することができる。
【００６１】
以下、作成手順を示す。
（１）１桁目の文字毎の数と総数を求める。（総数が少ない場合は、＊で代用する。）
（２）文字種毎の数を総数で割り、割合を求める。（１桁目の場合、「１」が８９％、「２」が１１％）
（３）１桁目が同じ文字のグループに関して、２桁目の文字毎の数と総数を求め、割合を求める。（１桁目「１」の場合、図１６の１〜５行目を集計し割合「０」６７％、「１」１８％、「２」１５％を得る）
（４）（３）の処理を２〜７桁まで繰り返す。（ｎ桁目の処理をするときには、ｎ−１桁目までの前文字列が等しいグループ毎に集計を行う。）
この結果、図１７に示すような、１〜７桁目までの出現率を得ることができる。
【００６２】
これにより、たとえば１、２桁目に対する、図５に示すような、出現率評価テーブル４１が総合評価部３５に登録される。
尚、出現率評価テーブル４１は、区分機１が安定稼働し稼働情報データが十分に畜積（通常１ヶ月程度）できた段階で作成する。
【００６３】
上記したように、区分機設置局で収集された郵便番号データの統計情報を利用して、文字の頻度（出現率）を反映した文字評価値を決定する。今まで人手で行っていた文字評価値決定のパラメータ調整を統計情報によって自動的に行う。
【００６４】
すなわち、複合類似度認識および構造特徴マッチング認識の各結果は、それぞれ総合評価部に送られ、ここで入力文字に対する答えと総合評価値を決定する。総合評価部では、郵便局にて収集された郵便番号データの統計情報により自動的に作成された出現率評価テーブル（出現率評価表）を用いて評価値の算出をする。すなわち、出現頻度の高い数字は評価値を高く、出現頻度の低い数字は評価値を低く出力するように、パラメータが設定されている。総合評価部の認識結果は、郵便番号編集部に送られ、ここで７桁の最終的な郵便番号候補が出力される。
【００６５】
これにより、複合類似度認識法と構造特徴マッチング法の２種類の文字認識結果から最終的な文字評価値を算出するときに使用するパラメータを人手によらず自動的に調整することができる。
【００６６】
また、区分機設置局で収集された郵便番号データの統計情報を利用してパラメータ調整を自動化することによって、大幅な省力化が図れる。
さらに、シミュレーションや過去の調整経験から得られる既知の情報をテーブルに反映させることによって、文字や辞書の特性を利用した細かい調整も可能となる。
【００６７】
複数の郵便番号候補を編集する処理において、全ての前文字列に対して総合評価処理（ステップ２６の処理）を行うことによって、全桁認識処理の計算量を削減することができる。
【００６８】
【発明の効果】
以上詳述したように、この発明によれば、区分装置で収集される数字列からなる所在情報の統計情報を利用してパラメータ調整を自動化することによって、大幅な省力化が図れる所在情報認識装置および区分装置を提供できる。
【図面の簡単な説明】
【図１】この発明の実施形態を説明するための区分機の概略構成を示す外観図。
【図２】区分機の概略構成を示す図。
【図３】区分機の概略構成を示すブロック図。
【図４】郵便番号認識部の概略構成を示すブロック図。
【図５】出現率評価テーブルの構成例を示す図。
【図６】結合テーブルの構成例を示す図。
【図７】レベル選択テーブルの構成例を示す図。
【図８】レベルテーブルの構成例を示す図。
【図９】自動調整処理を説明するためのフローチャート。
【図１０】出現率を求める際の処理を説明するためのフローチャート。
【図１１】調整パターン番号を求める際の処理を説明するためのフローチャート。
【図１２】頻度レベルを選択する際の処理を説明するためのフローチャート。
【図１３】総合評価値を算出する際の処理を説明するためのフローチャート。
【図１４】３桁目までの各桁ごとの総合評価点を説明するための図。
【図１５】３桁目までの各桁ごとの総合評価点（候補削減）を説明するための図。
【図１６】稼働情報ファイルの保存内容を説明するための図。
【図１７】１〜７桁目までの出現率を説明するための図。
【符号の説明】
１…区分機、Ｐ…郵便物、１５…郵便番号認識部、３１…複合類似度認識部、３１ａ…複合類似度辞書、３２…構造特徴マッチング認識部、３２ａ…構造特徴マッチング辞書、３３…複合類似度評価部、３４…構造特徴マッチング評価部、３５…総合評価部、４１…出現率評価テーブル、４２…結合テーブル、４３…レベル選択テーブル、４４…レベルテーブル。[0001]
BACKGROUND OF THE INVENTION
  The present invention provides a location information recognition device that reads location information consisting of a numeric string on a paper sheet and recognizes each character of the read numeric string.and,The present invention relates to a sorting apparatus that executes sorting processing using the recognition result of the location information recognition apparatus.
[0002]
[Prior art]
There are two types of character recognition methods (composite similarity method and structural feature matching method) in a mail address automatic reading and sorting machine that optically reads address information on mail as paper sheets and sorts the mail. in use.
[0003]
From the recognition results obtained by the two types of recognition methods, an evaluation value is obtained by using a parameter set in advance according to the appearance rate of the zip code (location information consisting of a numeric string), and the recognition result is comprehensively evaluated. For example, in each digit, the evaluation value of a number having a high appearance rate is set to be high, the value having a low appearance rate is set to a low value, and the like is set as a parameter.
[0004]
By performing such processing, numbers with a high appearance rate can be easily recognized, numbers with a low appearance rate can be made difficult to recognize, and the processing efficiency of the sorting machine in each region can be improved.
[0005]
For example, in post offices near Kawasaki, there are many mail items such as “21 *”. Therefore, by setting the first digit “2” and the second digit “1” to be easier to recognize than other numbers, the classification rate is improved and the misclassification rate is reduced. .
[0006]
Conventionally, this parameter has been adjusted based on the appearance rate of postal codes in each region, as experienced by local maintenance personnel.
As described above, parameter adjustment was done manually in each region, so it took a lot of time and effort, and there were many places that depended on the experience and intuition of people. It was.
[0007]
[Problems to be solved by the invention]
  The present invention relates to a location information recognition device that can achieve significant labor savings by automating parameter adjustment using statistical information of location information consisting of numeric strings collected by a sorting device.andThe object is to provide a sorting device.
[0008]
[Means for Solving the Problems]
  Of this inventionThe location information recognition device is a combination of reading means for reading location information composed of a numeric string on a paper sheet, and a plurality of character patterns for each character for reference for each character of the numeric string read by the reading means. A first output means for outputting a recognition candidate character, a type and reliability of the reference character pattern by a similarity method, and a character string read by the reading means for each character for reference Second output means for outputting a recognition candidate character and the type of the structural feature for reference by a structural feature matching method with a plurality of structural features, and storing the appearance rate of the character for each digit of the digit string The appearance rate is based on the first storage means associated with the digit before the number string, the output result of the first output means, and the output result of the second output means, Characters that are likely to be misread, The second storage means for storing the parameter number for adjustment with respect to the appearance rate of the misread character and the appearance rate of the recognized character, and the appearance rate of the recognition candidate character based on the output result of the second output means A third storage means for storing various reference values for determining a plurality of stages of frequency levels; a fourth storage means for storing an evaluation value for each frequency level and reliability for the parameter number for adjustment; The first reading means for reading out the appearance rate for each recognition candidate character by the first and second output means from the first storage means, the recognition candidate character by the first output means, and the reference Based on the type of character pattern, the candidate character recognized by the second output means, the type of structural feature for reference, and the appearance rate for each recognition candidate character read by the first read means Second reading means for reading the parameter number from the second storage means, the appearance rate for the recognition candidate characters by the second output means, the recognition candidate characters by the second output means, and the reference structure By comparing each reference value stored in the third storage unit against the feature type, a first determination unit that determines a frequency level and an adjustment for adjustment by the second reading unit Third reading means for reading the evaluation value stored in the fourth storage means based on the parameter number and the frequency level determined by the first determining means, and the first and second output means The second judgment means for judging the character of the number string based on the evaluation value read by the third reading means for each recognition candidate character according to the above, and the character of the number string judged by the second judgment means Collect to collect And means for changing the appearance rate of characters for each digit of the number string stored in the first storage means based on the collected contents of the characters of the number string collected by the collecting means; It has.
[0009]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 is an external view showing a sorting machine 1 according to the present invention, and FIG. 2 is a diagram showing a schematic configuration of the sorting machine 1. This sorter 1 has a large box-shaped sorter body 1a. This sorter 1 reads the address and postal code on the postal matter P, recognizes the address and postal code from the read contents, and sorts the postal matter P into the sorting destination corresponding to the recognized address and postal code. It is.
The sorting machine main body 1 a is provided with a supply unit 2, a scanner unit 3, a transport unit 4, a sorting unit 5, and a storage unit 6. When the postal matter P from the supply unit 2 is conveyed by the conveyance path R, it is guided to the storage unit 6 through the conveyance unit 4 and the sorting unit 5 in order.
[0010]
The supply unit 2 includes a placing table 7 on which the postal matter P is placed, and a taking-out unit 8 that takes out the postal matter P from the placing stand 7 one by one and sends it to the transport path R. The scanner unit 3 optically reads the entire image on each mail piece P conveyed by the conveyance path R and generates image information. The transport unit 4 transports the postal matter P that has passed through the scanner unit 3 to the sorting unit 5. The storage unit 6 has a large number of storage pockets 6a for sorting and storing the postal matter P. The sorting unit 5 sorts the postal matter P sent from the transport unit 4 into one of the storage pockets 6a,... Based on a recognition result described later with respect to image information from the scanner unit 3.
[0011]
The scanner unit 3 optically scans the postal matter P and photoelectrically converts it to obtain a pattern signal. For example, the scanner unit 3 receives a light source that irradiates light on the postal matter P, and the reflected light. It is composed of a self-scanning CCD image sensor or the like that converts signals. The output of the scanner unit 3 is supplied to the recognition unit 10. The recognizing unit 10 recognizes an address composed of a zip code, an address, and the like by recognizing a character pattern according to the output of the scanner unit 3.
[0012]
In the sorting machine 1, the supply unit 2, the scanner unit 3, the transport unit 4, the sorting unit 5, and the recognition unit 10 are connected to a control unit 11. The control unit 11 controls the overall operation of the sorting machine 1. For example, the control unit 11 reads out the segment designation data corresponding to the recognition result of the recognition unit 10 using the segment designation table stored in a memory (not shown), and reads the segment designation data (the storage pockets 6a,. The postal matter P is transported to the storage pockets 6a,.
[0013]
Further, the control section 11 controls the entire transport system by driving a transport mechanism section (not shown) such as the transport path R by a driver (not shown).
As shown in FIG. 3, the recognition unit 10 includes an area detection unit 12, a character line detection unit 13, a character detection unit 14, a zip code recognition unit 15, a character recognition unit 16, and a block character recognition unit 17. And an address recognition unit 18, a character dictionary unit 19, and an address dictionary unit 20.
[0014]
The area detection unit 12 detects an area in which an address and a zip code are described based on the shape and position of the area from the image supplied from the scanner unit 3.
The character line detection unit 13 detects a character line in which an address and a zip code in the area detected by the area detection unit 12 are described. The address line is detected from the image data in the area detected by the area detecting unit 12.
[0015]
The character line detection unit 13 detects a stamp, a postal code frame, an advertisement text, etc. from the image in the region detected by the region detection unit 12, removes them, and then detects a line from the arrangement of characters. A priority is given to a line that seems to be an address or a zip code line based on the size and arrangement of the line, and the detection result is output in order from the candidate with the highest priority.
[0016]
The character detection unit 14 detects and cuts out characters of the character line detected by the character line detection unit 13 one character at a time, that is, detects and cuts out character information corresponding to an address (postal code).
The zip code recognizing unit 15 stores character data (numerical data: standard pattern corresponding to numbers) stored in the character dictionary unit 19 with the characters (postal codes) supplied from the character detecting unit 14, that is, detected and cut out. And the postal code is recognized. The zip code recognition method and the recognition means in the zip code recognition unit 15 will be described in detail later.
[0017]
The character recognizing unit 16 recognizes a character by collating the character supplied from the character detecting unit 14, that is, the detected cut-out character with the character data stored in the character dictionary unit 19. Similarly, the block character recognizing unit 17 recognizes the characters of the block by comparing the character supplied from the character detecting unit 14, that is, the detected cut-out character with the character data stored in the character dictionary unit 19. The address recognizing unit 18 recognizes the address by comparing the character supplied from the character detecting unit 14, that is, the detected cut-out character with the address data stored in the address dictionary unit 20.
[0018]
The zip code recognition result from the zip code recognition unit 15, the character recognition result from the character recognition unit 16, the block character recognition result from the block character recognition unit 17, and the address recognition result from the address recognition unit 18 are: It is supplied to the control unit 11. The control unit 11 discriminates an address according to each recognition result supplied, and outputs classification designation data for the determined address, that is, data indicating the storage pockets 6a in the sorting unit 5.
[0019]
FIG. 4 shows the configuration of the zip code recognition unit 15 in detail. That is, the character patterns of the character detection cutout candidates supplied from the character detection unit 14 are sent to the composite similarity recognition unit 31 and the structural feature matching recognition unit 32, respectively. The composite similarity recognition unit 31 performs a recognition process by performing a composite similarity calculation by a comparison operation between the supplied character pattern and the standard pattern of the composite similarity dictionary 31a, and the recognition result is the composite similarity of the upper candidate. Are left alone and sent to the composite similarity evaluation unit 33. Since the composite similarity dictionary 31a has a plurality of entry characters for one character type, the entry similarity has different similarities. The composite similarity evaluation unit 33 calculates an evaluation value called similarity reliability based on a threshold value statistically predicted in advance from the similarity distribution, and recognizes up to the top three candidates of the calculated evaluation value. Determine as a result.
[0020]
The structural feature matching recognition unit 32 calculates the contour structure of the supplied character pattern, and decomposes the contour structure feature into character components called segments. The structural feature matching recognizing unit 32 recognizes the structure from the connection condition of the segment, and the segment is composed of direction code and length information. The structural feature matching recognizing unit 32 performs a transition of the structural feature matching dictionary 32a, and the structural feature matching is sent to the structural feature matching evaluation unit 34. The structure feature matching dictionary 32a has a plurality of entry characters for one character type, each of which reflects the shape of the character. Information about which entry character matches the structural feature is also sent to the structural feature matching evaluation unit 34 at the same time. The structural feature matching evaluation unit 34 re-evaluates the detailed characteristics of the characters and determines up to three recognition results.
[0021]
Each recognition result of the composite similarity evaluation unit 33 and the structural feature matching evaluation unit 34 is sent to the comprehensive evaluation unit 35, and an answer and a comprehensive evaluation value for the character pattern supplied by the comprehensive evaluation unit 35 are determined. The comprehensive evaluation unit 35 obtains the appearance rate for each digit by using the appearance rate evaluation table (first storage unit) 41 shown in FIG. 5, and combines the table (second storage unit) shown in FIG. 42, the adjustment pattern number is selected, the frequency selection level is selected using the level selection table (third storage means) 43 shown in FIG. 7, and the level table (fourth storage) shown in FIG. Means) 44 is used to calculate the evaluation value. (Such evaluation value output method using statistical information is hereinafter referred to as automatic adjustment processing.)
The evaluation result of the comprehensive evaluation unit 35 is sent to the postal code editing unit 36, where a final 7-digit postal code candidate is output.
[0022]
As shown in FIG. 5, the appearance rate evaluation table 41 includes the appearance rate of each character for each digit. For example, the appearance rate of 1 is 89% and the appearance rate of 2 is 11 as the appearance rate of the first digit. %, The appearance rate of other numbers is 0%. As the appearance rate of the second digit when the first digit is 1, the appearance rate of 0 is 67%, the appearance rate of 1 is 18%, the appearance rate of 2 is 15%, and the appearance rate of other numbers is 0 %. As the appearance rate of the second digit when the first digit is 2, the appearance rate of 0 is 100%, and the appearance rate of other numbers is 0%.
[0023]
The appearance rate evaluation table 41 is created from statistical information of postal code data collected in advance, as will be described later.
As shown in FIG. 6, the connection table 42 includes characters (recognition results) for structural feature matching, entries (distinguished by differences in the shapes of characters to be compared for matching), and characters with complex similarity (recognition results). And entries (distinguished by differences in character patterns to be compared for composite similarity), characters that are likely to be misread (other characters), and other characters' conditions (the appearance rate of the other characters) ), An adjustment pattern number for the condition (appearance rate of recognized character−appearance rate of partner character = appearance rate difference) is registered.
[0024]
For example, when the structural feature matching character is “0”, the entry is “1”, the composite similarity character is “0”, the entry is “1”, and the partner character is “6”, the appearance rate is the partner condition. When the appearance rate difference is larger than “6” and the appearance rate difference is larger than “6”, the adjustment pattern number “5” is selected.
[0025]
In addition, when the structural feature matching character is “0”, the entry is “1”, the composite similarity character is “0”, the entry is “1”, and the partner character is “8”, the appearance rate is the partner condition. When it is larger than “5”, the adjustment pattern number “2” is selected.
[0026]
In addition, when the structural feature matching character is “0”, the entry is “1”, the composite similarity character is “0”, the entry is “1”, and the partner character is “8”, the appearance rate is the partner condition. When the appearance rate difference is smaller than “6” and the appearance rate difference is larger than “20”, the adjustment pattern number “3” is selected.
[0027]
Also, when the structural feature matching character is “0”, the entry is “1”, the composite similarity character is “0”, the entry is “1”, and the partner character, the partner's condition, and your condition are other than the above In addition, the adjustment pattern number “10” is selected.
Further, when the structural feature matching character is “0”, the entry is “2”, the composite similarity character is “0”, the entry is “other than *: 1”, and the partner character is “6”, the partner condition is The adjustment pattern number “1” is selected when the appearance rate is smaller than “6” and the difference in appearance rate is smaller than “10” as a condition of the user.
[0028]
As shown in FIG. 7, the level selection table 43 defines five levels of frequency levels of strong strengthening, strengthening, standard, weakening, and weakening depending on the appearance rate of the recognition result of the structural feature matching. In this level selection table 43, a threshold value between levels is set for each entry of structural feature matching characters (recognition result) (identified by the difference in the shape of characters to be compared for matching).
[0029]
For example, when the character of the structural feature matching is “0” and the entry is “0”, the threshold value for distinguishing between strengthening and strengthening is “20”, the threshold value for distinguishing between strengthening and standard is “10”, and standard and weakening The threshold for distinguishing between “5” and “2” for distinguishing between weakening and weakening is “2”.
[0030]
In addition, when the structural feature matching character is “0” and the entry is “1”, the threshold value for distinguishing between strengthening and strengthening is “25”, the threshold value for distinguishing between strengthening and standard is “15”, and standard and weakening. The threshold for distinguishing between “10” and “5” for distinguishing between weakening and weakening is “10”.
In the level table 44, as shown in FIG. 8, evaluation values are registered based on the adjustment pattern number, the frequency level, and the reliability of the composite similarity.
[0031]
For example, for the adjustment pattern number “1”, the frequency level is strongly enhanced and the reliability of the composite similarity is “0”, “1”, “2”, “3”, respectively, and the evaluation values “2”, “ 4 ”,“ 4 ”,“ 5 ”are stored correspondingly, the frequency level of the adjustment pattern number“ 1 ”is strengthened, and the reliability of the composite similarity is“ 0 ”,“ 1 ”,“ 2 ”,“ 3 ”, evaluation values“ 2 ”,“ 3 ”,“ 4 ”,“ 5 ”are stored in correspondence with each other, and the frequency level of the adjustment pattern number“ 1 ”is standard and the reliability of the composite similarity is standard. Are stored in correspondence with the evaluation values “1”, “3”, “4” and “5” for “0”, “1”, “2” and “3”, respectively. ”For the evaluation values“ 1 ”,“ 3 ”,“ 4 ”,“ 4 ”, respectively, with the frequency level weakened and the reliability of the composite similarity is“ 0 ”,“ 1 ”,“ 2 ”,“ 3 ”. 4 " And the evaluation value “0”, “1”, “2”, “3” for the adjustment pattern number “1”, the frequency level is weakened, and the reliability of the composite similarity is “0”. “1”, “2”, “3”, “4” are stored correspondingly.
[0032]
Also, for the adjustment pattern number “2”, the frequency level is strongly enhanced and the reliability of the composite similarity is “0”, “1”, “2”, “3”, respectively, with evaluation values “2”, “ “4”, “4”, “5” are stored correspondingly, the frequency level of the adjustment pattern number “2” is strengthened, and the reliability of the composite similarity is “0”, “1”, “2”, “ 3 ”, evaluation values“ 2 ”,“ 3 ”,“ 5 ”,“ 5 ”are stored in correspondence with each other, and the frequency of the adjustment pattern number“ 2 ”is standard and the reliability of the composite similarity is standard. Are stored in correspondence with the evaluation values “2”, “3”, “4” and “5” for “0”, “1”, “2” and “3”, respectively. ”For the evaluation values“ 1 ”,“ 3 ”,“ 4 ”,“ 4 ”, respectively, with the frequency level weakened and the reliability of the composite similarity is“ 0 ”,“ 1 ”,“ 2 ”,“ 3 ”. 5 ”corresponds And the evaluation value “1” for the adjustment pattern number “2” with the frequency level weakening and the reliability of the composite similarity being “0”, “1”, “2”, “3”, respectively. , “3”, “4”, “4” are stored correspondingly.
[0033]
For the adjustment pattern number “3”, the frequency level is strongly enhanced and the reliability of the composite similarity is “0”, “1”, “2”, “3”, respectively, and the evaluation values “3”, “3”, respectively. “5”, “6”, “6” are stored correspondingly, the frequency level of the adjustment pattern number “3” is strengthened, and the reliability of the composite similarity is “0”, “1”, “2”, “ 3 ”, evaluation values“ 3 ”,“ 4 ”,“ 6 ”,“ 6 ”are stored in correspondence with each other, and the frequency of the adjustment pattern number“ 3 ”is standard and the reliability of the composite similarity is standard. Are stored in correspondence with the evaluation values “3”, “4”, “5”, “6” for “0”, “1”, “2”, “3”, respectively. ”, And the reliability of the composite similarity is“ 0 ”,“ 1 ”,“ 2 ”,“ 3 ”, the evaluation values“ 2 ”,“ 4 ”,“ 5 ”,“ 6 ”corresponds And the evaluation value “2” for the adjustment pattern number “3” with the frequency level weakening and the reliability of the composite similarity being “0”, “1”, “2”, “3”, respectively. , “4”, “5”, “5” are stored correspondingly.
[0034]
Next, automatic adjustment processing in the above configuration will be described with reference to the flowchart shown in FIG.
For example, as a specific example, a procedure for obtaining the evaluation value of the second digit “0” of the postal code “100-0001” by automatic adjustment processing is shown. Assume that the following data is given as input data. For simplicity, in the following, both the recognition result of the composite similarity and the recognition result of the structural feature matching are described as “0-1” when the character is “0” and the entry is 1, for example.
[0035]

In the comprehensive evaluation, evaluation values are calculated for all combinations of the recognition result of the composite similarity and the recognition result of the structural feature matching, and the top two candidates are used as answers to the previous character string. Specifically, in the case of (Example 1), the previous character string is “1” and “2” respectively.
(Composite similarity recognition result, structural feature matching recognition result)
= (0-1, 0-1), (0-1, 8-1), (6-2, 0-1),
(6-2, 8-1), (0-1, x), (6-1, x), (x, 0-1),
(×, 8-1)
For all combinations (× is a state in which there are no candidates), the overall evaluation values are calculated by performing the operations in the following 22nd to 25th steps, and one candidate or two candidates ( For the details of selecting the number of candidates, refer to the 26th step) as the answer to the previous character string candidate.
[0036]
In the twenty-second to twenty-fifth steps, the previous character string is “1” and (recognition result of composite similarity, recognition result of structural feature matching) = (0-1, 0-1) as an example. I will give you a description.
[0037]
In a twenty-second step, the character appearance rate is obtained. This will be described with reference to FIG. In the appearance rate evaluation table 41, the appearance rate of characters at each digit is set in a manner depending on the previous character string candidate. FIG. 5 is an example of the appearance rate evaluation table 41 for the first and second digits. In this case, the probability of “0” appearing in the second digit is 67% when the first digit is “1”, but is 100% when the first digit is “2”. Indicates. Similarly, when the first digit is “1”, the probability that “1” appears in the second digit is 18%, the probability that “2” appears is 15%, and the others are 0%. In the automatic adjustment process, when the appearance rate of a certain digit is obtained, the appearance rate evaluation table 41 as an appearance rate evaluation table is searched using the previous character string and the recognition result of the structural feature matching of the currently evaluated digit.
[0038]
In the case of (Example 1), the first digit candidate is “1”, and the second digit structural feature matching candidates are “0” and “8”. Therefore, an appearance rate of 67% of “0” is obtained.
Next, how to obtain the appearance rate when the digit to be processed is the N-th digit (1 ≦ N ≦ 7), the recognition result of the structural feature matching is CN, the previous character string C1, C2... CN-1 = Cn, This will be described with reference to the flowchart shown in FIG.
[0039]
That is, when the N-digit appearance rate is obtained, n is set to 1 (ST31), and whether n matches N (ST32) is checked. The appearance rate (A1) of the feature matching recognition result CN is obtained (ST35).
[0040]
If they do not match as a result of the check in step 32, a pointer to the next digit is obtained from the appearance rate evaluation value of the first digit (ST33), 1 is added to n (n ++) (ST34), and the process returns to step 32.
As a result of the check in step 32, if they match, the process proceeds to step 35, where the appearance rate (A1) of the recognition result CN of the structural feature matching is obtained from the N-th digit appearance rate evaluation table 41. Proceed to 33.
[0041]
In the 23rd step, an adjustment pattern number is obtained as shown in the flowchart of FIG.
From the connection table 42, a line in which the recognition result of the structural feature matching and the recognition result of the composite similarity match is searched (ST41, 42, 50), and the “partner condition” and the “own condition” are satisfied (ST43˜). 47) The pattern number of the first line (I) is output as the adjustment pattern number (A2) (ST48).
[0042]
When the last line is reached, the pattern number of the last line is output as the adjustment pattern number (A2) (ST49).
The “partner character” is a number of the partner that is assumed to be easily misread, and is set from known information obtained from simulations and past adjustment experiences. For example, information such as a pattern in which a result of 0-1 in the structural feature matching recognition result is often misread as “6” is included in the combination table 42.
[0043]
In the case of the example, the recognition result of the structural feature matching is 0-1, and the recognition result of the composite similarity is 0-1, and at this time, the first to fourth lines in FIG. Search whether the conditions match in order.
First line:
From the appearance rate evaluation table 41 as the appearance rate evaluation table, the appearance rate of “6” is 0% in the current digit, which does not match the “partner's condition”.
[0044]
Second line:
Similarly, the appearance rate of “8” is 0%, and “the other party's condition” does not match.
3rd line:
The other party's condition matches with (0 <6), and his condition matches with ((67-0)> 20), and this line is selected. As a result, in this example, adjustment pattern number 3 is selected.
[0045]
In the twenty-fourth step, a frequency level is selected.
Depending on the appearance rate of the recognition result of the structural feature matching, five levels of frequency levels of strong strengthening, strengthening, standard, weakening and weakening are determined. In the level selection table 43, a threshold value between levels is set for each entry of the recognition result of the structural feature matching. Since the structure feature matching entry reflects the character characteristics, the frequency level can be changed depending on the reliability of the entry.
[0046]
Now, when the level selection table 43 shown in FIG. 7 is given, in the case of (Example 1), the recognition result of structural feature matching is 0-1 and the appearance rate is 67% (> 25%) in the 23rd step. This frequency level is a strong enhancement.
[0047]
Next, frequency level selection will be described with reference to the flowchart shown in FIG.
That is, the characters and entries that match the recognition result of the structural feature matching are sequentially checked from the first line of the level selection table 43 (ST51, 52, 54, and 55). If the frequency level is higher than the threshold for distinguishing the frequency level of reinforcement (ST56), the frequency level (A3) is determined as strong reinforcement (ST57), and the appearance rate is greater than the threshold for distinguishing between the frequency level of reinforcement and the standard. (ST58) The frequency level (A3) is determined to be strengthened (ST59), and if the appearance rate is greater than the threshold for distinguishing between standard and weakening (ST60), the frequency level (A3) is determined to be standard (ST61) When the rate is larger than the threshold value for discriminating between weakening and weakening (ST62), the frequency level (A3) is determined as weakening (ST63), and the appearance rate is a threshold for discriminating between weakening and weakening It determines that the weak attenuating (ST62) frequency level (A3) is smaller than (ST64).
[0048]
In addition, if there is a mismatch even if the last line is checked in step 52, the frequency level (A3) is standardized as an exception process (ST53).
In the twenty-fifth step, a comprehensive evaluation value is calculated.
[0049]
An evaluation value is selected from the level table 44 using the adjustment pattern number and frequency level obtained in the 23rd and 24th steps, and the reliability of the composite similarity.
When the level table 44 shown in FIG. 8 is used, in the case of (Example 1), the recognition result of the composite similarity is 0-1 (3), the pattern number 3, and the frequency level is strongly enhanced. Becomes “6”.
[0050]
Next, calculation of the comprehensive evaluation value will be described with reference to the flowchart shown in FIG.
In other words, those that match the adjustment pattern number (A2) obtained in the 23rd step are sequentially checked from the first row of the level table 44 (ST71, 72, 73, 76). A comprehensive evaluation value (A4) is selected and output from the frequency level (A3) obtained in the step and the reliability of the composite similarity (ST74).
[0051]
In addition, if there is no match even if the last line (I) is checked in step 72, the comprehensive evaluation value “0” (A4 = 0) is output as an exception process (ST75).
[0052]
In a twenty-sixth step, comprehensive evaluation processing is performed for all previous character strings. Even if the recognition result of the same composite similarity and the recognition result of structural feature matching are input, the overall evaluation value may change depending on the previous character string candidate.
In the case of (Example 1), after evaluating the combination of the recognition results of all the composite similarities and the recognition results of structural feature matching for the previous character string “1”, the same applies to the case of the previous character string “2”. Comprehensive evaluation is performed (ST27 to ST29, ST22 to ST26).
[0053]
As a result, when the comprehensive evaluation is performed up to the seventh digit (ST27), a plurality of postal code candidates are obtained. The comprehensive evaluation value obtained for each digit is added for each candidate, and this is used as the comprehensive evaluation point of the postal code candidate (ST13). The top two candidates with a high overall evaluation score are selected as final results (ST14).
[0054]
If the overall evaluation calculation operation as described above is simply performed and processing is performed by selecting from the highest overall evaluation score to the second highest in each digit for the existing character string, for example, up to the third digit, FIG. (The number in parentheses is the overall evaluation score). As a result, in the comprehensive evaluation of the fourth digit, there are eight types of previous character string candidates.
[0055]
When the above operations are performed up to 7 digits, 128 candidates are obtained, and the total number of times of comprehensive evaluation for each digit is 127 times, and the combined similarity recognition result is 127 times respectively. All the combinations of the recognition results of the structural feature matching are evaluated, and the amount of calculation becomes enormous. Therefore, the amount of calculation can be reduced by the following rules.
[0056]
Select up to 2 candidates for the first digit.
In the second and subsequent digits, two candidates are selected only from the binary tree shown in FIG. 14 that are connected to the node with the higher overall evaluation score, and only one candidate is selected at the other nodes.
When the comprehensive evaluation up to the third digit is performed by this operation, the result is as shown in FIG. 15, and the comprehensive evaluation is performed on the four previous character string candidates at the fourth digit. When processing up to the seventh digit by this method, final candidates are narrowed down to eight candidates. Among the eight candidates, the one with the highest total sum of evaluation points for each digit is selected up to the second place in order, and is set as the final zip code candidate.
[0057]
In the above embodiment, the case where the character recognition method is applied to two recognition methods of the composite similarity method and the structural feature matching method has been described. However, the present invention is not limited to this. Alternatively, the present invention can be similarly applied when both recognition methods are replaced with another method. Further, the present invention can be similarly applied when there is one kind of recognition method.
[0058]
Furthermore, the tables to be used are not limited to this. In the above embodiment, the method for reflecting the known misread information obtained by experience or simulation in the table has been described, but in order to reflect the frequency more simply, for example, obtained from the recognition result. A method is also conceivable in which the similarity is multiplied by the appearance rate to obtain a comprehensive evaluation value.
[0059]
Next, how to determine the appearance rate will be described.
The appearance rate evaluation table 41 as an appearance rate evaluation table is a file in which the appearance frequency of 0 to 9 in each digit is set in a form depending on the previous sentence order, and the operation information file 11a as shown in FIG. Create from. The postal codes that exist in the control unit 11 and are actually processed (recognized) by the sorting machine 1 and the number of mails are tabulated and stored.
[0060]
The operation information file 11a is present in a device that manages a plurality of sorting machines for each post office, and is obtained by collecting and storing the postal codes actually processed (recognized) by each sorting machine and the number of mails. Also good.
Conventionally, the appearance rate of characters that the operator has empirically captured is replaced with statistical information obtained from an actual machine, and can be more accurately reflected in the character recognition process.
[0061]
The creation procedure is shown below.
(1) Find the number and total number of characters in the first digit. (If the total number is small, substitute with *.)
(2) Divide the number for each character type by the total number to obtain the ratio. (In the first digit, "1" is 89%, "2" is 11%)
(3) With respect to a group of characters having the same first digit, the number and total number of characters in the second digit are obtained, and the ratio is obtained. (In the case of “1” in the first digit, the 1st to 5th lines in FIG. 16 are totaled to obtain ratios “0” 67%, “1” 18%, “2” 15%)
(4) The process of (3) is repeated up to 2 to 7 digits. (When the n-th digit is processed, the calculation is performed for each group in which the previous character strings up to the (n-1) -th digit are equal.)
As a result, the appearance rate up to the first to seventh digits as shown in FIG. 17 can be obtained.
[0062]
Thereby, for example, the appearance rate evaluation table 41 as shown in FIG. 5 for the first and second digits is registered in the comprehensive evaluation unit 35.
The appearance rate evaluation table 41 is created when the sorting machine 1 is stably operated and the operation information data is sufficiently stocked (usually about one month).
[0063]
As described above, the character evaluation value reflecting the character frequency (appearance rate) is determined using the statistical information of the zip code data collected by the sorting machine installation station. The parameter adjustment for character evaluation value determination, which has been performed manually until now, is automatically performed by statistical information.
[0064]
That is, each result of the composite similarity recognition and the structural feature matching recognition is sent to a comprehensive evaluation unit, where an answer to the input character and a comprehensive evaluation value are determined. The comprehensive evaluation unit calculates an evaluation value using an appearance rate evaluation table (appearance rate evaluation table) automatically created based on statistical information of postal code data collected at a post office. In other words, the parameters are set so that numbers with high appearance frequency output high evaluation values and numbers with low appearance frequency output low evaluation values. The recognition result of the comprehensive evaluation unit is sent to the postal code editing unit, where a final 7-digit postal code candidate is output.
[0065]
Thereby, it is possible to automatically adjust the parameters used when calculating the final character evaluation value from the two types of character recognition results of the composite similarity recognition method and the structural feature matching method, regardless of the human hand.
[0066]
In addition, automating parameter adjustment using statistical information of postal code data collected by the sorting machine installation station can save a lot of labor.
Further, by reflecting known information obtained from simulations and past adjustment experiences in the table, fine adjustments using the characteristics of characters and dictionaries are possible.
[0067]
In the process of editing a plurality of postal code candidates, the total evaluation process (the process of step 26) is performed on all the previous character strings, thereby reducing the calculation amount of the all-digit recognition process.
[0068]
【The invention's effect】
  As described above in detail, according to the present invention, a location information recognition device that can achieve significant labor savings by automating parameter adjustment using statistical information of location information consisting of numeric strings collected by a sorting device.andA sorting device can be provided.
[Brief description of the drawings]
FIG. 1 is an external view showing a schematic configuration of a sorting machine for explaining an embodiment of the present invention.
FIG. 2 is a diagram showing a schematic configuration of a sorting machine.
FIG. 3 is a block diagram showing a schematic configuration of a sorting machine.
FIG. 4 is a block diagram showing a schematic configuration of a postal code recognition unit.
FIG. 5 is a diagram showing a configuration example of an appearance rate evaluation table.
FIG. 6 is a diagram showing a configuration example of a join table.
FIG. 7 is a diagram showing a configuration example of a level selection table.
FIG. 8 is a diagram showing a configuration example of a level table.
FIG. 9 is a flowchart for explaining automatic adjustment processing;
FIG. 10 is a flowchart for explaining processing when obtaining an appearance rate;
FIG. 11 is a flowchart for explaining processing when obtaining an adjustment pattern number;
FIG. 12 is a flowchart for explaining processing when a frequency level is selected.
FIG. 13 is a flowchart for explaining processing when calculating a comprehensive evaluation value;
FIG. 14 is a diagram for explaining a comprehensive evaluation score for each digit up to the third digit.
FIG. 15 is a diagram for explaining a comprehensive evaluation score (candidate reduction) for each digit up to the third digit;
FIG. 16 is a diagram for explaining the saved contents of an operation information file.
FIG. 17 is a diagram for explaining the appearance rate up to the first to seventh digits.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... Sorting machine, P ... Mail, 15 ... Postal code recognition part, 31 ... Composite similarity recognition part, 31a ... Composite similarity dictionary, 32 ... Structural feature matching recognition part, 32a ... Structural feature matching dictionary, 33 ... Composite Similarity evaluation unit, 34 ... structural feature matching evaluation unit, 35 ... comprehensive evaluation unit, 41 ... appearance rate evaluation table, 42 ... combination table, 43 ... level selection table, 44 ... level table.

Claims (3)

Reading means for reading location information consisting of a string of numbers on a paper sheet;
The recognition candidate character, the type of the reference character pattern, and the reliability are output by a composite similarity method with a plurality of character patterns for each character of each character string read by the reading means. First output means;
A second recognition character and a type of structural feature for reference are output by a structural feature matching method with a plurality of structural features for each character of each digit string read by the reading means. Output means;
Storing an appearance rate of characters for each digit of the digit string , wherein the appearance rate is associated with a digit of a digit before the digit sequence ;
Based on the output result of the first output means and the output result of the second output means, adjustment for the characters that are likely to be misread, the appearance rate of the misread characters, and the appearance rate of the recognized characters Second storage means for storing the parameter number of
Third storage means for storing various reference values for determining a plurality of frequency levels for the appearance rate of recognition candidate characters based on the output result of the second output means;
Fourth storage means for storing an evaluation value for each frequency level and reliability for the parameter number for adjustment;
First reading means for reading an appearance rate for each recognition candidate character by the first and second output means from the first storage means;
The recognition candidate character by the first output unit and the type of the reference character pattern, the recognition candidate character by the second output unit and the type of the structural feature for reference are read by the first reading unit. Second reading means for reading the parameter number for adjustment from the second storage means based on the appearance rate for each recognition candidate character;
Each of the appearance rates for the recognition candidate characters by the second output means and the recognition candidate characters by the second output means and the types of structural features for reference are stored in the third storage means. A first determination means for determining a frequency level by comparing with a reference value;
Third reading means for reading the evaluation value stored in the fourth storage means based on the parameter number for adjustment by the second reading means and the frequency level determined by the first determination means; ,
Second judging means for judging a character of the number string based on an evaluation value read by the third reading means for each recognition candidate character by the first and second output means;
A collecting means for collecting the characters of the numeric string determined by the second determining means;
Based on the collected contents of the characters of the numeric string collected by the collecting means, changing means for changing the appearance rate of the characters for each digit of the numeric string stored in the first storage means;
A location information recognition apparatus characterized by comprising: Reading means for reading location information consisting of a string of numbers on a paper sheet;
The recognition candidate character, the type of the reference character pattern, and the reliability are output by a composite similarity method with a plurality of character patterns for each character of each character string read by the reading means. First output means;
A second recognition character and a type of structural feature for reference are output by a structural feature matching method with a plurality of structural features for each character of each digit string read by the reading means. Output means;
First determination means for determining characters of the numeric string based on the output result of the first output means and the output result of the second output means;
A collecting means for collecting the characters of the numeric string determined by the first determining means;
First storage means for storing an appearance rate of characters for each digit of the numeric string based on the collected contents of the characters of the numeric string collected by the collecting means;
Based on the output result of the first output means and the output result of the second output means, adjustment for the characters that are likely to be misread, the appearance rate of the misread characters, and the appearance rate of the recognized characters Second storage means for storing the parameter number of
Third storage means for storing various reference values for determining a plurality of frequency levels for the appearance rate of recognition candidate characters based on the output result of the second output means;
Fourth storage means for storing an evaluation value for each frequency level with respect to the parameter number for adjustment;
First reading means for reading an appearance rate for each recognition candidate character by the first and second output means from the first storage means;
The recognition candidate character by the first output unit and the type of the reference character pattern, the recognition candidate character by the second output unit and the type of the structural feature for reference are read by the first reading unit. Second reading means for reading the parameter number for adjustment from the second storage means based on the appearance rate for each recognition candidate character;
Each of the appearance rates for the recognition candidate characters by the second output means and the recognition candidate characters by the second output means and the types of structural features for reference are stored in the third storage means. A second determination means for determining a frequency level by comparing with a reference value;
Third reading means for reading the evaluation value stored in the fourth storage means based on the parameter number for adjustment by the second reading means and the frequency level determined by the second determination means; ,
Third judging means for judging the character of the numeric string based on the evaluation value read by the third reading means for each recognition candidate character by the first and second output means;
A location information recognition apparatus characterized by comprising: Recognizing location information consisting of a number string on a paper sheet using the location information recognition apparatus according to claim 1 or 2, and sorting the paper sheet based on the recognition result. Sorting equipment to do.

Images