JP3628312B2 - Watermark information embedding device and watermark information detection device - Google Patents

Description

【００６９】
ここで剰余をＲｎ（Ｒｎ＝Ｃｎ−（Ｐｗ×（Ｐｈ−１）））とすると，ユニットパターン行列にはＤｎ回のデータ符号ユニットおよびデータ符号の先頭Ｒｎビット分に相当するユニットパターンを埋め込むことになる。ただし，剰余部分のＲｎビットは必ずしも埋め込まなくても良い。
【００７０】
図９の説明では，ユニットパターン行列のサイズを９×１１（１１行９列），データ符号長を１２（図中で０〜１１の番号がついたものがデータ符号の各符号語を表わす）とする。
【００７１】
次いで，ユニットパターン行列の第１行目に符号長データを埋め込む（ステップＳ２０３）。図９の例では符号長を９ビットのデータで表現して１度だけ埋め込んでいる例を説明しているが，ユニットパターン行列の幅Ｐｗが十分大きい場合，データ符号と同様に符号長データを繰り返し埋め込むこともできる。
【００７２】
さらに，ユニットパターン行列の第２行以降に，データ符号ユニットを繰り返し埋め込む（ステップＳ２０４）。図９で示すようにデータ符号のＭＳＢ（ｍｏｓｔ ｓｉｇｎｉｆｉｃａｎｔ ｂｉｔ）またはＬＳＢ（ｌｅａｓｔ ｓｉｇｎｉｆｉｃａｎｔ ｂｉｔ）から順に行方向に埋め込む。図９の例ではデータ符号ユニットを７回，およびデータ符号の先頭６ビットを埋め込んでいる例を示している。
【００７３】
データの埋め込み方法は図９のように行方向に連続になるように埋め込んでも良いし，列方向に連続になるように埋め込んでも良い。
【００７４】
以上，透かし画像形成部１２における，透かし画像について説明した。次いで，透かし情報埋め込み装置１０の透かし入り文書画像合成部１３について説明する。
【００７５】
（透かし入り文書画像合成部１３）
透かし入り文書画像合成部１３では，文書画像形成部１１で作成した文書画像と，透かし画像形成部で作成した透かし画像を重ね合わせる。透かし入り文書画像の各画素の値は，文書画像と透かし画像の対応する画素値の論理積演算（ＡＮＤ）によって計算する。すなわち，文書画像と透かし画像のどちらかが０（黒）であれば，透かし入り文書画像の画素値は０（黒），それ以外は１（白）となる。
【００７６】
図１０は，透かし入り文書画像の一例を示す説明図である。図１１は，図１０の一部を拡大して示した説明図である。ここで，ユニットパターンは図７（１）のパターンを用いている。透かし入り文書画像は，出力デバイス１４により出力される。
【００７７】
以上，透かし情報埋め込み装置１０の動作について説明した。
次いで，図１，及び，図１２〜図２０を参照しながら，透かし情報検出装置３０の動作について説明する。
【００７８】
（透かし検出部３２）
図１２は，透かし検出部３２の処理の流れを示す流れ図である。
まず，スキャナなどの入力デバイス３１によって透かし入り文書画像を計算機のメモリ等に入力する（ステップＳ３０１）。この画像を入力画像と称する。入力画像は多値画像であり，以下では２５６階調のグレイ画像として説明する。また入力画像の解像度（入力デバイス３１で読み込むときの解像度）は，上記透かし情報埋め込み装置１０で作成した透かし入り文書画像と異なっていても良いが，ここでは上記透かし情報埋め込み装置１０で作成した画像と同じ解像度であるとして説明する。また，入力画像は回転や伸縮などの補正が行われているものとする。
【００７９】
次いで，入力画像の大きさと信号ユニットの大きさから，ユニットパターンがいくつ埋め込まれているかを計算する（ステップＳ３０２）。例えば入力画像の大きさがＷ（幅）×Ｈ（高さ）であるとして，信号ユニットの大きさをＳｗ×Ｓｈ，ユニットパターンはＵｗ×Ｕｈ個のユニットから構成されるとすると，入力画像中に埋め込まれているユニットパターンの数（Ｎ＝Ｐｗ×Ｐｈ）は以下のように計算される。
【００８０】
【数２】

【００８１】
ただし，透かし情報埋め込み装置１０と透かし情報検出装置３０で解像度が異なる場合には，それらの解像度の比によって入力画像中の信号ユニットの大きさを正規化した後，上記の計算を行う。
【００８２】
次いで，ステップＳ３０２で計算したユニットパターン数をもとに入力画像に対してユニットパターンの区切り位置を設定する（ステップＳ３０３）。図１３は入力画像（図１３（１））と，ユニットパターンの区切り位置を設定した後の入力画像（図１３（２））の一例を示している。
【００８３】
次いで，ユニットパターンの区切りごとにシンボルユニットの検出を行い，ユニットパターン行列を復元する（ステップＳ３０４）。以下に，信号検出の詳細を説明する。
【００８４】
図１４は，入力画像中における，図３（１）に示したユニットＡに対応する領域の一例を示した説明図である。図３では信号ユニットは二値画像であるが，ここでは多値画像である。二値画像を印刷した場合，インクのにじみなどが原因で濃淡が連続的に変化するため，図１４のようにドットの周囲が白と黒の中間色になる。したがって図１４を波の伝播方向と平行な方向から見た断面図は図１５のようになる。図４が矩形波であるのに対し，図１５は滑らかな波となる。
【００８５】
また，実際には紙の厚さの局所的な変化や，印刷文書の汚れ，出力デバイスや画像入力デバイスの不安定性などの要因により，入力画像中には多くの雑音成分が付加されることになるが，ここでは雑音成分のない場合について説明する。しかしながら，ここで説明する方法を用いれば，雑音成分が付加された画像に対しても安定した信号検出を行うことができる。
【００８６】
以下では入力画像から信号ユニットを検出するために，波の周波数と方向，および影響範囲を同時に定義できる二次元ウェーブレットフィルタを用いる。以下では，二次元ウェーブレットフィルタの一つであるガボールフィルタを用いる例を示すが，ガボールフィルタと同様な性質を持つフィルタであれば，必ずしもガボールフィルタである必要はなく，さらには信号ユニットと同じドットパターンを持つテンプレートを定義してパターンマッチングを行うなどの方法でも良い。
【００８７】
以下にガボールフィルタＧ（ｘ，ｙ），ｘ＝０〜ｇｗ−１，ｙ＝０〜ｇｈ−１を示す。ｇｗ，ｇｈはフィルタのサイズであり，ここでは上記透かし情報埋め込み装置１０で埋め込んだ信号ユニットと同じ大きさである。
【００８８】
【数３】

【００８９】
信号検出には透かし画像に埋め込んだシンボルユニットと周波数，波の方向，および大きさが等しいガボールフィルタを，埋め込んだ信号ユニットの種類と同じ数だけ用意する。ここでは図３のユニットＡとユニットＢに対応するガボールフィルタをフィルタＡ，フィルタＢと称する。
【００９０】
入力画像中の任意の位置でのフィルタ出力値はフィルタと画像間のコンボリューションにより計算する。ガボールフィルタの場合は実数フィルタと虚数フィルタ（虚数フィルタは実数フィルタと半波長分位相がずれたフィルタ）が存在するため，それらの２乗平均値をフィルタ出力値とする。例えば，フィルタＡの実数フィルタと画像間のコンボリューションがＲｃ，虚数フィルタとのコンボリューションがＩｃであったとすると，出力値Ｆ（Ａ）は以下の式で計算する。
【００９１】
【数４】

【００９２】
図１６は，ステップＳ３０３によって区切られたユニットパターンＵ（ｘ，ｙ）中に埋め込まれているシンボルユニットがユニットＡであるかユニットＢであるかを判定する方法について説明する説明図である。
【００９３】
ユニットパターンＵ（ｘ，ｙ）に対するシンボル判定ステップを以下のように行う。
（１）フィルタＡの位置を移動しながら，ユニットパターンＵ（ｘ，ｙ）中のすべての位置についてＦ（Ａ）を計算した結果の最大値をユニットパターンＵ（ｘ，ｙ）に対するフィルタＡの出力値とし，これをＦｕ（Ａ，ｘ，ｙ）とする。
（２）ユニットパターンＵ（ｘ，ｙ）に対するフィルタＢの出力値を（１）と同様に計算し，これをＦｕ（Ｂ，ｘ，ｙ）とする。
（３）Ｆｕ（Ａ，ｘ，ｙ）とＦｕ（Ｂ，ｘ，ｙ）を比較し，Ｆｕ（Ａ，ｘ，ｙ）≧Ｆｕ（Ｂ，ｘ，ｙ）であればユニットパターンＵ（ｘ，ｙ）に埋め込まれているシンボルユニットはユニットＡであると判定し，Ｆｕ（Ａ，ｘ，ｙ）＜Ｆｕ（Ｂ，ｘ，ｙ）であればユニットパターンＵ（ｘ，ｙ）に埋め込まれているシンボルユニットはユニットＢであると判定する。
【００９４】
（１），（２）において，フィルタを移動するステップ幅は任意であり，ユニットパターン上の代表的な位置での出力値のみを計算してもよい。また，（３）でＦｕ（Ａ，ｘ，ｙ）とＦｕ（Ｂ，ｘ，ｙ）の差の絶対値があらかじめ定めておいた閾値以下であった場合には判定不能としてもよい。
【００９５】
また（１）において，フィルタをずらしながらコンボリューションを計算する過程で，Ｆ（Ａ）の最大値があらかじめ定めた閾値を超えた場合に，ただちにＵ（ｘ，ｙ）に埋め込まれているシンボルユニットはユニットＡであると判定して処理を中止しても良い。（２）においても同様に，Ｆ（Ｂ）の最大値があらかじめ定めた閾値を超えた場合に，ただちにＵ（ｘ，ｙ）に埋め込まれているシンボルユニットはユニットＢであると判定しても良い。
【００９６】
以上，信号検出（ステップＳ３０４）の詳細について説明した。再び，図１２の流れ図に戻り，以降のステップＳ３０５について説明する。ステップＳ３０５では，ユニットパターン行列のシンボルを連結してデータ符号を再構成し，元の情報を復元する。
【００９７】
図１７は情報復元の一例を示す説明図である。情報復元のステップは以下の通りである。
（１）各ユニットパターンに埋め込まれているシンボルを検出する（図１７▲１▼）。
（２）シンボルを連結してデータ符号を復元する（図１７▲２▼）。
（３）データ符号を復号して埋め込まれた情報を取り出す（図１７▲３▼）。
【００９８】
図１８〜図２０はデータ符号の復元方法の一例を示す説明図である。復元方法は基本的に図８の逆の処理となる。
【００９９】
まず，ユニットパターン行列の第１行から符号長データ部分を取り出して，埋め込まれたデータ符号の符号長を得る（ステップＳ４０１）。
【０１００】
次いで，ユニットパターン行列のサイズとＳ４０１で得たデータ符号の符号長をもとに，データ符号ユニットを埋め込んだ回数Ｄｎおよび剰余Ｒｎを計算する（ステップＳ４０２）。
【０１０１】
次いで，ユニットパターン行列の２行目以降からステップＳ２０３と逆の方法でデータ符号ユニットを取り出す（ステップＳ４０３）。図１９の例ではＵ（１，２）（２行１列）から順に１２個のパターンユニットごとに分解する（Ｕ（１，２）〜Ｕ（３，３），Ｕ（４，３）〜Ｕ（６，４），・・・）。Ｄｎ＝７，Ｒｎ＝６であるため，１２個のパターンユニット（データ符号ユニット）は７回取り出され，剰余として６個（データ符号ユニットの上位６個に相当する）のユニットパターン（Ｕ（４，１１）〜Ｕ（９，１１））が取り出される。
【０１０２】
次いで，ステップＳ４０３で取り出したデータ符号ユニットに対してビット確信度演算を行うことにより，埋め込んだデータ符号を再構成する（ステップＳ４０４）。以下，ビット確信度演算について説明する。
【０１０３】
図２０のようにユニットパターン行列の２行１列目から最初に取り出されたデ−外符号ユニットをＤｕ（１，１）〜Ｄｕ（１２，１）とし，順次Ｄｕ（１，２）〜Ｄｕ（１２，２），・・・，と表記する。また，剰余部分はＤｕ（１，８）〜Ｄｕ（６，８）とする。ビット確信度演算は各データ符号ユニットの要素ごとに多数決を取るなどして，データ符号の各シンボルの値を決定することである。これにより，文字領域との重なりや紙面の汚れなどが原因で，任意のデータ符号ユニット中の任意のユニットから正しく信号検出を行えなかった場合（ビット反転エラーなど）でも，最終的に正しくデータ符号を復元することができる。
【０１０４】
具体的には例えばデータ符号の１ビット目は，Ｄｕ（１，１），Ｄｕ（１，２），・・・，Ｄｕ（１，８）の信号検出結果が１である方が多い場合には１と判定し，０である方が多い場合には０と判定する。同様にデータ符号の２ビット目はＤｕ（２，１），Ｄｕ（２，２），・・・，Ｄｕ（２，８）の信号検出結果による多数決によって判定し，データ符号の１２ビット目はＤｕ（１２，１），Ｄｕ（１２，２），・・・，Ｄｕ（１２，７）（Ｄｕ（１２，８）は存在しないためＤｕ（１２，７）までとなる）の信号検出結果による多数決によって判定する。
【０１０５】
ビット確信度演算は，図１６の信号検出フィルタの出力値を加算することによっても行うこともできる。これは，例えば図３（１）のユニットＡに０のシンボルが割り当てられ，図３（２）のユニットＢに１のシンボルが割り当てられているものとし，Ｄｕ（ｍ，ｎ）に対するフィルタＡによる出力値の最大値をＤｆ（Ａ，ｍ，ｎ），Ｄｕ（ｍ，ｎ）に対するフィルタＢによる出力値の最大値をＤｆ（Ｂ，ｍ，ｎ）とすると，データ符号のＭビット目は，
【０１０６】
【数５】

【０１０７】
の場合は１と判定し，それ以外の場合は０と判定する。ただし，Ｎ＜Ｒｎの場合はＤｆの加算はｎ＝１〜Ｒｎ＋１までとなる。
【０１０８】
ここではデータ符号を繰り返し埋め込む場合について説明したが，データを符号化する際に誤り訂正符号などを用いることにより，データ符号ユニットの繰り返しを行わないような方法も実現できる。
【０１０９】
以上詳細に説明したように，本実施の形態によれば，以下のような優れた効果がある。
（１−１）ドットの配列の違いにより埋め込み情報を表現するため，元の文書のフォント，文字間や行間のピッチに対する変更を伴わない。
（１−２）シンボルを割り当てているドットパターンと，シンボルを割り当てていないドットパターンの濃度（一定区間内のドットの数）が等しいため，人の目には文書の背景に一様な濃度の網掛けがされているように見え，情報の存在が目立たない。
（１−３）シンボルを割り当てているドットパターンと割り当てていないドットパターンを秘密にしておくことで，埋め込まれている情報の解読が困難となる。
（１−４）情報を表わすパターンは細かいドットの集まりで，文書の背景として一面に埋め込まれているため，埋め込みアルゴリズムが公開されたとしても，印刷された文書に対する埋め込み情報の改ざんが困難となる。
（１−５）波（濃淡変化）の方向の違いにより埋め込み信号を検出するため（１画素単位の詳細な検出を行わないので），印刷文書に多少の汚れなどがあった場合でも，安定した情報検出を行うことができる。
（１−６）同じ情報を繰り返し埋め込み，検出時には繰り返し埋め込まれた情報のすべてを利用して情報復元を行うため，大きなフォントの文字によって信号部分が隠されたり，用紙が汚れていたりすることによる部分的な情報の欠落が発生しても，安定して埋め込んだ情報を取り出すことができる。
【０１１０】
（第２の実施の形態）
上記第１の実施の形態の文書出力部では，透かしとして印刷物に挿入するデータを誤り訂正符号などを用いて符号語（この符号語は暗号化されていてもよい）に変換する。例えば２元符号に変換した場合，０，１のビット列でデータが表現される。各シンボル（シンボル０，シンボル１）にはそれを表現するドットパターンが対応する。これが文書画像の背景として埋め込まれるが，１枚の紙に埋め込むことが可能な最大信号数に比べて符号語のビット数（ドットパターン配列の要素数）が小さい場合，ドットパターンの配列は，紙面上に繰り返し埋め込まれることになる。
【０１１１】
透かし信号は文書画像の背景として埋め込まれるため，信号を表わすドットパターンの一部，または全部が文書画像の文字領域と重なっている場所からは，信号検出時にドットパターンを検出できない。上記第１の実施の形態で説明しているように，信号検出時において，符号語の各ビットが０であるか１であるかの判定は，各ビットに対応する位置から検出されたドットパターンがどちらであるかの多数決などの方法による。従って，ドットパターンの検出が不可能となる位置（文字領域と重なった領域）が符号語のあるビットに集中したような場合に，そのビットは，図２１に示したように，判定不可能になる。符号語に誤り訂正符号を用いたとしても，判定不可能なビットが誤り訂正能力を超えた場合には情報を復元できない。
【０１１２】
図２２は，本実施の形態にかかる透かし情報埋め込み装置及び透かし情報検出装置の構成を示す説明図である。上記第１の実施の形態との相違点は，透かし情報埋め込み装置４０に，埋め込みパターン判定部４２が追加されたことである。埋め込みパターン判定部４２は，紙面上にどのような順序で透かし信号を配置するかを決定する部分である。
【０１１３】
その他の構成要素については，上記第１の実施の形態と同様であるため重複説明を省略することとし，本実施の形態の動作について，上記第１の実施の形態への追加事項を中心に説明する。
【０１１４】
図２３は，埋め込みパターン判定部４２，および透かし画像形成部１２の処理の流れを示している。破線で囲まれた領域（ステップＳ５０１〜Ｓ５０５）が埋め込みパターン判定部４２に相当する部分である。
【０１１５】
埋め込みパターン判定部４２の概要は，あらかじめＮｐ種類の信号埋め込みパターンを用意しておき，実際に信号の配置を行う前に，仮にそれぞれの埋め込みパターンに従って埋め込んだ場合に，文字と重なる場所がどの程度の割合で特定のビットに集中するかどうかを判断し，文字と重なる部分が符号語のすべてのビットに最も広く散らばるような埋め込みパターンを選択することである。
【０１１６】
符号長をＬ，文書中に埋め込むことができる信号（ドットパターン）数のうち文字と重ならない信号数をＮとしたとき，符号語の各ビットの有効埋め込み回数（文字と重ならずに埋め込むことができる回数）の平均値ＶはＮ／Ｌである。したがって，理想的には符号語のすべてのビットがＶ回ずつ文字と重ならない領域に配置することができる埋め込みパターンが最も良いパターンになる。前述したように，有効埋め込み回数に偏りがあると，少ない回数しか埋め込めなかったビットは，信号検出時に信号検出誤り，印刷後の紙面の汚れ，紙面上への新たな書き込みなどにより信号の読み取りが不可能になる可能性が高い。偏りが大きい場合には，汚れや書き込みによっても全く影響を受けないビットが存在する一方で，信号読み取り不可能なビットが数多く出現することになる。したがって，符号語のすべてのビットが有効埋め込み回数の平均値に近い値である，すなわち各ビットの有効埋め込み回数の分散（または標準偏差）が最も小さいパターンを選択すればよい。
【０１１７】
ステップＳ５０１では透かしとして印刷物に挿入するデータを誤り訂正符号などを用いて符号語（この符号語は暗号化されていてもよい）に変換する。例えば２元符号に変換した場合，０，１のビット列でデータが表現される。各シンボル（シンボル０，シンボル１）にはそれを表現するドットパターンが対応する。
【０１１８】
ステップＳ５０２〜Ｓ５０４ではＮｐ種類の埋め込みパターンについて（ステップＳ５０２），埋め込もうとする符号語の各ビットが文字領域と重なる場合の数を，各ビット毎にカウントする（ステップＳ５０３）。
【０１１９】
図２４は４種類の埋め込みパターンの例である（Ｎｐ＝４）。図２４（ａ）は水平方向に連続して信号を埋め込み，（ｂ）は垂直方向に連続，（ｃ），（ｄ）は斜め方向に連続して信号を埋め込んでいる。これ以外にも，渦巻き型に埋め込むなど，無数に考えられる。さらに，図２５のように文書画像をいくつかのブロックに分割し，それぞれのブロックで異なる埋め込みパターンを用いても良い。
【０１２０】
図２６は符号語の例とそれに基づいてドットパターンを配列した例である。このドットパターンを繰り返し文書画像の背景として配置する。
【０１２１】
図２７（ａ）は図２４（ａ）の埋め込みパターンに従って図２６の符号語を配置した例を示している。ここでは図２６のｍビット目のドットパターンは１２回埋め込まれている。図２７（ｂ）はドットパターンの前景に印刷文字を重ねあわせたものである。図のように，符号語のｍビット目は１回目の配置では文字と重なっていないが，２回目の配置では文字と重なっている。図２４の埋め込みパターンＰに対して，符号語のｍビット目が文字と重ならずに配置できた場合の数をＮ（Ｐ，ｍ）と表記する。符号語の各ビットの有効埋め込み回数の総和Ｔ（Ｐ）はＮ（Ｐ，ｍ）をｍ＝１〜Ｌまで加算したものであり，平均有効埋め込み回数Ｅ（Ｐ）はＴ（Ｐ）／Ｌとなる。これより，有効埋め込み回数の分散は，
Ｖ（Ｐ）＝Ｅ｛Ｅ（Ｐ）−Ｎ（Ｐ，ｍ））^２｝
で計算できる。
【０１２２】
すべての埋め込みパターンについてＶ（Ｐ）を計算し，ステップＳ５０５によりＶ（Ｐ）が最も小さくなる埋め込みパターンを当該の文書画像に対する埋め込みパターンであると確定する。
【０１２３】
ステップＳ５０６ではステップＳ５０５で確定した埋め込みパターンに従って透かし信号を配置し，文書画像と重ねあわせる。このとき図２８で示すようにヘッダ情報領域にどの埋め込みパターンで埋め込んだかという情報を同時に埋め込む。ヘッダ情報領域は文書によらず一定の埋め込みパターンで情報が配置されているものとする。
【０１２４】
次いで，透かし検出部３２について説明する。
図２９は透かし検出部３２の処理の流れを示している。
ステップＳ６０１では入力画像のうちヘッダ情報が埋め込まれている部分の信号を検出し，ヘッダ情報を復元する。
ステップＳ６０２ではステップＳ６０１で復元したヘッダ情報から信号埋め込みパターンを取り出す。
ステップＳ６０３ではステップＳ６０２で取り出した埋め込みパターンに従って，透かし信号を読み取る。
ステップＳ６０４では符号語の各ビット毎に多数決をとり，各ビットの値を確定し，符号を複合して情報を取り出す。
【０１２５】
なお，図２３のステップＳ５０６では埋め込みパターンをヘッダ情報埋め込み領域に記録したが，埋め込みパターンが公開されており，かつ数種類に限定されるならば埋め込みパターンの記録をせずに，検出時にすべてのパターンで情報の復号を行って，正しく復号できた（復号エラーの起きない，または意味のある情報が取り出せた）情報を透かし情報として取り出すことも可能である。
【０１２６】
また，埋め込みパターンを秘密にする場合も埋め込みパターンの記録を行う必要はない。これは透かし情報が秘密の情報であった場合などに，埋め込みパターンをより複雑にして，どのようなパターンがあるかを公開せず，信号を埋め込んだパターンを秘密にしておくことで，より安全性を保つことを目的としている。このような目的で利用する場合は，必ずしもステップＳ５０２〜Ｓ５０５の処理を行う必要はない。
【０１２７】
また，埋め込みパターンのブロック分割（図２５）についての補説すると，信号埋め込みパターンの１つとして文書画像をいくつかのブロックに分割し，それぞれのブロックで異なる埋め込みパターンを適用する場合に，ブロックの分割方法としては図２５に示すように固定の大きさのブロックで分割する以外に，以下で説明するような方法を用いても良い。
【０１２８】
前提として，例えば横書きの文書に水平方向に連続した埋め込みパターン（図２４（ａ））を採用した場合は，文字と重なる領域がある程度連続するため，符号長によっては，図２１に示したような判定不可能の状況が起こりやすい。縦書きの文書に垂直方向（図２４（ｂ））に連続した埋め込みパターンを採用した場合も同様である。
【０１２９】
図３０は縦書き，横書き，図などが混在した文書の例である。このような文書では図２４のような文書全体に対して１つの埋め込みパターンを適用するよりも，いくつかのブロックに分割して，それぞれに異なる埋め込みパターンを適用する方が，符号語の各ビットの有効埋め込み回数の分散値がより小さく押さえられる。分割方法は先に説明したような固定の大きさのブロックに分割しても良いが，文書画像の特徴を解析して動的な分割を行った方が分散値低減の効果が大きくなる。
【０１３０】
図３１（ａ）は図３０の文書を縦書き，横書き等の特徴によって分割した例である。この分割はＯＣＲの前処理として用いられる「ＯＣＲの表解析（特開平０４−０３３０７９）」などの方法を用いてもよいし，図３０の文書画像を縮小して，ある程度の大きさの文字領域のかたまりを抽出し，そのかたまりを１つのブロックとして分割しても良い。図３１（ａ）の例では，領域１，領域２，領域４は横書きの文字列の領域，領域５は縦書きの文字列の領域，領域３は図の領域となっている。
【０１３１】
これらの各領域毎に図２３のステップＳ５０２〜Ｓ５０５を行い，各領域の埋め込みパターンを決定する。図３１（ｂ）は領域毎に最適な埋め込みパターンが適用された例である。
【０１３２】
文書画像がどのように分割されたかという情報と，各領域に対する埋め込みパターンの情報は図２８のヘッダ情報領域に記録しても良いし，秘密の情報としても良い。
【０１３３】
また，文書画像中の図領域には信号を埋め込むことが困難なため，図３１（ａ）で領域分割した際に，図領域と判定された領域には信号を埋め込まず，図領域に信号を埋め込まなかったという情報をヘッダ情報領域に記録するようにしてもよい。図領域に信号を埋め込まないことによって，透かし検出時に元々信号がほとんど埋め込まれていない領域から信号検出を行うという処理を省略することができるため，処理時間の短縮や信号検出誤りを防ぐことができる。
【０１３４】
以上詳細に説明したように，本実施の形態によれば，以下のような優れた効果がある。
（２−１）紙の背景に微小なドットパターンによって視覚的に違和感のない方法で情報を埋め込む場合において，前景の文字領域とドットパターンの重なりによって検出付可能となる信号が，情報を表わす符号語の一部のビットに集中することがないので，信号検出時の検出誤りや，紙の汚れなどの劣化，紙に上書がなされた場合でも確実に情報を取り出すことができる。
（２−２）信号の埋め込みパターンを秘密にしておくことで，埋め込みパターンを知っているものだけが情報を取り出すことができるようになり，埋め込んだ情報を悪意を持つ第三者に盗み見られる危険性が減少する。
（２−３）文書画像を構成要素の特徴毎に領域分割し，分割された領域毎に最適な信号埋め込みパターンを適用することで，（２−１）の効果をより高めることができる。
（２−４）文書画像を構成要素の特徴毎に領域分割し，図領域など信号をほとんど埋め込むことができない部分には信号を埋め込まないことで，透かし検出時の処理時間を短縮し，また検出誤りを低減できる。
【０１３５】
以上，添付図面を参照しながら本発明にかかる透かし情報埋め込み装置，及び，透かし情報検出装置の好適な実施形態について説明したが，本発明はかかる例に限定されない。当業者であれば，特許請求の範囲に記載された技術的思想の範疇内において各種の変更例または修正例に想到し得ることは明らかであり，それらについても当然に本発明の技術的範囲に属するものと了解される。
【０１３６】
【発明の効果】
以上説明したように，本発明によれば，以下のような優れた効果がある。
（１−１）ドットの配列の違いにより埋め込み情報を表現するため，元の文書のフォント，文字間や行間のピッチに対する変更を伴わない。
（１−２）シンボルを割り当てているドットパターンと，シンボルを割り当てていないドットパターンの濃度（一定区間内のドットの数）が等しいため，人の目には文書の背景に一様な濃度の網掛けがされているように見え，情報の存在が目立たない。
（１−３）シンボルを割り当てているドットパターンと割り当てていないドットパターンを秘密にしておくことで，埋め込まれている情報の解読が困難となる。
（１−４）情報を表わすパターンは細かいドットの集まりで，文書の背景として一面に埋め込まれているため，埋め込みアルゴリズムが公開されたとしても，印刷された文書に対する埋め込み情報の改ざんが困難となる。
（１−５）波（濃淡変化）の方向の違いにより埋め込み信号を検出するため（１画素単位の詳細な検出を行わないので），印刷文書に多少の汚れなどがあった場合でも，安定した情報検出を行うことができる。
（１−６）同じ情報を繰り返し埋め込み，検出時には繰り返し埋め込まれた情報のすべてを利用して情報復元を行うため，大きなフォントの文字によって信号部分が隠されたり，用紙が汚れていたりすることによる部分的な情報の欠落が発生しても，安定して埋め込んだ情報を取り出すことができる。
【０１３７】
また，本発明によれば，以下のような優れた効果がある。
（２−１）紙の背景に微小なドットパターンによって視覚的に違和感のない方法で情報を埋め込む場合において，前景の文字領域とドットパターンの重なりによって検出付可能となる信号が，情報を表わす符号語の一部のビットに集中することがないので，信号検出時の検出誤りや，紙の汚れなどの劣化，紙に上書がなされた場合でも確実に情報を取り出すことができる。
（２−２）信号の埋め込みパターンを秘密にしておくことで，埋め込みパターンを知っているものだけが情報を取り出すことができるようになり，埋め込んだ情報を悪意を持つ第三者に盗み見られる危険性が減少する。
（２−３）文書画像を構成要素の特徴毎に領域分割し，分割された領域毎に最適な信号埋め込みパターンを適用することで，（２−１）の効果をより高めることができる。
（２−４）文書画像を構成要素の特徴毎に領域分割し，図領域など信号をほとんど埋め込むことができない部分には信号を埋め込まないことで，透かし検出時の処理時間を短縮し，また検出誤りを低減できる。
【図面の簡単な説明】
【図１】第１の実施の形態にかかる透かし情報埋め込み装置及び透かし情報検出装置の構成を示す説明図である。
【図２】透かし画像形成部１２の処理の流れを示す流れ図である。
【図３】透かし信号の一例を示す説明図であり，（１）はユニットＡを，（２）はユニットＢを示している。
【図４】図３（１）の画素値の変化をａｒｃｔａｎ（１／３）の方向から見た断面図である。
【図５】透かし信号の一例を示す説明図であり，（３）はユニットＣを，（４）はユニットＤを，（５）はユニットＥを示している。
【図６】背景画像の説明図であり，（１）はユニットＥを背景ユニットと定義し，これを隙間なく並べた透かし画像の背景とした場合を示し，（２）は（１）の背景画像の中にユニットＡを埋め込んだ一例を示し，（３）は（１）の背景画像の中にユニットＢを埋め込んだ一例を示している。
【図７】透かし画像へのシンボル埋め込み方法の一例を示す説明図である。
【図８】機密情報を透かし画像に埋め込む方法について示した流れ図である。
【図９】透かし検出部３２の処理の流れを示す流れ図である。
【図１０】透かし入り文書画像の一例を示す説明図である。
【図１１】図１０の一部を拡大して示した説明図である。
【図１２】透かし検出部３２の処理の流れを示す流れ図である。
【図１３】図１３は入力画像（図１３（１））と，ユニットパターンの区切り位置を設定した後の入力画像（図１３（２））の一例を示している。
【図１４】入力画像中におけるユニットＡに対応する領域の一例を示した説明図である。
【図１５】図１４を波の伝播方向と平行な方向から見た断面図である。
【図１６】ユニットパターンＵ（ｘ，ｙ）中に埋め込まれているシンボルユニットがユニットＡであるかユニットＢであるかを判定する方法について説明する説明図である。
【図１７】情報復元の一例を示す説明図である
【図１８】データ符号の復元方法の一例を示す説明図である。
【図１９】データ符号の復元方法の一例を示す説明図である。
【図２０】データ符号の復元方法の一例を示す説明図である。
【図２１】ドットパターンの検出が不可能になる場合を示す説明図である。
【図２２】第２の実施の形態にかかる透かし情報埋め込み装置及び透かし情報検出装置の構成を示す説明図である。
【図２３】埋め込みパターン判定部４２の処理の流れを示す流れ図である。
【図２４】埋め込みパターンの例を示す説明図であり，（ａ）は水平方向，（ｂ）は垂直方向，（ｃ）は右上から左下への斜め方向，（ｄ）は左上から右下への斜め方向を示している。
【図２５】文書画像を領域分割した場合を示す説明図である。
【図２６】符号語の例とそれに基づいてドットパターンを配列した例を示す説明図である。
【図２７】図２４（ａ）の埋め込みパターンに従って図６の符号語を配列した例を示す説明図である。
【図２８】ヘッダ情報領域と透かし情報領域を示す説明図である。
【図２９】透かし検出部３２の処理の流れを示す流れ図である。
【図３０】縦書き，横書き，図などが混在した文書の例を示す説明図である。
【図３１】図３０の文書を縦書き，横書き，図などの特徴によって分割した例を示す説明図である。
【符号の説明】
１０ 透かし情報埋め込み装置
１１ 文書画像形成部
１２ 透かし画像形成部
１３ 透かし入り文書画像合成部
１４ 出力デバイス
１５ 文書データ
１６ 機密情報
２０ 印刷文書
３０ 透かし情報検出装置
３１ 入力デバイス
３２ 透かし検出部
４０ 透かし情報埋め込み装置
４２ 埋め込みパターン判定部[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a method for adding confidential information to a document image in a format other than characters, and a technique for detecting confidential information from a printed document containing confidential information.
[0002]
[Prior art]
“Digital watermarking” that embeds information to prevent copying / counterfeiting and confidential information in images and document data in a form that is invisible to the human eye is based on the premise that all storage and data transfer are performed on electronic media. Since the information embedded by the watermark is not deteriorated or lost, the information can be reliably detected. Similarly, documents printed on paper media cannot be easily tampered with in a form that is not visually unsightly other than text to prevent unauthorized tampering and copying. There is a need for a method for embedding confidential information in a printed document.
[0003]
The following techniques are known as information embedding methods for black and white binary documents that are most widely used as printed materials.
[0004]
[1] Japanese Patent Laid-Open No. 2001-78006 “Method and apparatus for embedding / detecting watermark information in a monochrome binary document image”
A minimum rectangle surrounding an arbitrary character string is divided into several blocks and divided into two groups (group 1 and group 2) (the number of groups may be three or more). For example, when the signal is 1, the feature amount in the block of group 1 is increased and the feature amount in each block of group 2 is decreased. When the signal is 0, the reverse operation is performed. The feature amount in the block includes the number of pixels in the character area, the thickness of the character, and the distance to the point where the block first hits the character area.
[0005]
[2] JP 2001-53954 A "Information embedding device, information reading device, digital watermark system, information embedding method, information reading method, and recording medium"
The width and height of the minimum rectangle that encloses one character are defined as the feature amount for that character, and the symbol is represented by a classification pattern of the magnitude relationship of the feature amount between two or more characters. For example, six feature quantities can be defined from three characters, and combinations of patterns of these magnitude relationships are enumerated, these combinations are classified into two grooves, and a symbol is given to each. If the information to be embedded is “0” and the combination pattern of the feature amounts of the character selected to represent this is “1”, one of the six feature amounts is expanded in the character area, etc. To change. The pattern to be changed is selected so that the amount of change is minimized.
[0006]
[3] Japanese Patent Laid-Open No. 9-179494 “Confidential Information Recording Method”
Assume that printing is performed with a printer of 400 dpi or more. Information is digitized and information is expressed by the distance (number of dots) between the reference point mark and the position determination mark.
[0007]
[4] Japanese Patent Application No. 10-200743 “Document Processing Device”
Information is expressed by whether or not the screen lines of the line screen (special screen composed of fine parallel lines) are moved backward.
[0008]
[Problems to be solved by the invention]
However, in the known techniques [1] and [2], the font and layout are changed because the change is made to the pixels constituting the characters of the document image, the character spacing, and the line spacing. In addition, in the above known techniques [3] and [4], when detecting, precise detection processing for each pixel of an input image read from an input device such as a scanner is required. When noise is added at the time of reading, the information detection accuracy is greatly affected.
[0009]
As described above, in the known techniques [1] to [4], when the printed document is input to the computer again by an input device such as a scanner and the embedded confidential information is detected, the smear or input of the printed document is detected. There is a problem that it is difficult to accurately extract secret information because many noise components are included in the input image due to image deformation such as rotation that occurs during the process.
[0010]
The present invention has been made in view of the above-mentioned problems of the conventional watermark information embedding / detection technique, and an object of the present invention is to provide new and improved watermark information that can accurately extract secret information. To provide an embedding device and a watermark information detecting device.
[0011]
[Means for Solving the Problems]
In order to solve the above problems, according to a first aspect of the present invention, a watermark information embedding device is provided. The watermark information embedding device of the present invention includes: Dot pattern on image Superimpose Watermarked image Is provided with a watermarked image composition unit.
[0012]
And There are multiple types of dot patterns, It is regularly arranged, and at least one kind of dot pattern therein is characterized by being given confidential information representing a specific secret.
[0013]
In addition, by preparing a plurality of dot patterns in which the wave direction and / or wavelength is changed depending on the dot arrangement, giving one symbol to one dot pattern, and arranging the dot patterns in combination, You may comprise so that confidential information may be given.
[0014]
The symbol may be composed of a valid symbol that forms part of the confidential information and an invalid symbol that is irrelevant to the confidential information.
[0015]
Alternatively, a plurality of dot patterns whose wave directions and / or wavelengths are changed depending on the dot arrangement, one symbol is given to one dot pattern arrangement combination, and the dot pattern arrangement combination is changed. The confidential information can be given by arranging them in combination.
[0016]
In addition, Watermarked image The confidential information may be repeatedly embedded in an arbitrary range on the paper output by the output device or on the entire surface.
[0017]
In order to solve the above problem, according to a second aspect of the present invention, a watermark information detection apparatus is provided. The watermark information detection apparatus of the present invention is image And a watermark image in which multiple types of dot patterns are embedded. Watermarked image And a watermark detection unit for detecting the watermark image.
[0018]
The watermark detection unit includes a filter that extracts a plurality of types of dot patterns that are the same as a watermark image. Watermarked image On the other hand, the watermark image is detected by performing matching.
[0019]
In addition, printed on paper Watermarked image The watermark detection unit reads the input device. Watermarked image On the other hand, the watermark image may be detected by performing matching.
[0020]
In addition, the watermark image is prepared by preparing a plurality of dot patterns in which the wave direction and / or wavelength is changed according to the dot arrangement, giving one symbol to one dot pattern, and arranging the dot patterns in combination. Thus, confidential information is given, and the filter may be configured to include the same number of two-dimensional wavelet filters as the dot pattern having the same wave direction and wavelength as the dot pattern. . As an example of a two-dimensional wavelet filter, a Gabor filter can be used.
[0021]
In such a case, a convolution (convolution integration) between an arbitrary region in the watermarked image and the plurality of two-dimensional wavelet filters is calculated, and the dot pattern corresponding to the two-dimensional wavelet filter that maximizes the convolution is obtained. It is possible to determine that it is embedded in the area.
[0022]
In addition, when the convolution of an arbitrary region in the watermarked image and an arbitrary two-dimensional wavelet filter exceeds a certain threshold, the dot pattern corresponding to the two-dimensional wavelet filter is embedded in the region. It is also possible to judge that
[0023]
Further, when the confidential information is repeatedly embedded in an arbitrary range or on the entire surface of the paper input to the input device, all signal detection at positions corresponding to the same code bit in the watermarked image is performed. A convolution with a filter is added for each of the plurality of two-dimensional wavelet filters as many times as the number of repetitions at the time of embedding, and the dot pattern corresponding to the two-dimensional wavelet filter having the largest sum is embedded. It is possible to judge.
[0024]
According to another aspect of the present invention, the following watermark information embedding device is provided as an application example of the above-described watermark information embedding device. This watermark embedding device is characterized in that, in addition to the constituent elements of the watermark information embedding device, further includes a pattern determination unit that selects what kind of embedding pattern the plurality of types of dot patterns are arranged.
[0025]
The pattern determination unit changes the number of embedding patterns of the dot pattern (for example, 2 or more) and temporarily arranges the dot pattern before actually arranging the dot pattern, image The embedding pattern can be selected in accordance with the number of dot patterns that can be detected even if they are superimposed on the watermark image (hereinafter referred to as the number of effective dot patterns).
[0026]
More specifically, it is possible to calculate the variance of the number of effective dot patterns of each dot pattern for each of the embedded patterns subjected to the temporary arrangement, and select the embedded pattern having the smallest variance. It is also possible to calculate the standard deviation instead of the variance and select the embedding pattern having the smallest standard deviation.
[0027]
Information on the embedded pattern selected by the pattern determination unit is Watermarked image Can be embedded as header information. At this time, if the information related to the embedding pattern is encrypted in order to keep it secret to a third party, only the person who knows the embedding pattern can restore the embedded watermark information.
[0028]
Said image It is also possible to divide the area and select the optimum embedding pattern for each divided area. As an example of area division, image Then, an analysis process (characteristic recognition process) of characters and tables is performed using an OCR (Optical Character Reader). image It is possible to divide the area according to the characteristics.
[0029]
image As a result of the characteristic recognition process, there may be a case where there is an area (for example, an area shown in the figure) in which the dot pattern (watermark information) can be hardly embedded. It is possible not to arrange the dot pattern in such a divided area. Then, information on the divided areas where the dot pattern is not arranged is Watermarked image If it is embedded in the header information, it is not necessary to detect the watermark information from the divided area when detecting the watermark information, and the detection efficiency can be improved.
[0030]
DETAILED DESCRIPTION OF THE INVENTION
Exemplary embodiments of a watermark information embedding device and a watermark information detection device according to the present invention will be described below in detail with reference to the accompanying drawings. In the present specification and drawings, components having substantially the same functional configuration are denoted by the same reference numerals, and redundant description is omitted.
[0031]
(First embodiment)
FIG. 1 is an explanatory diagram showing the configuration of a watermark information embedding device and a watermark information detection device according to this embodiment.
[0032]
(Watermark information embedding device 10)
The watermark information embedding device 10 is a device that forms a document image based on document data and confidential information embedded in the document and prints it on a paper medium. As shown in FIG. 1, the watermark information embedding device 10 includes a document image forming unit 11, a watermark image forming unit 12, a watermarked document image synthesizing unit 13, and an output device 14. The document data 15 is data created by a document creation tool or the like. The confidential information 16 is information (character string, image, audio data) to be embedded in a paper medium in a format other than characters.
[0033]
In the document image forming unit 11, an image in a state where the document data 15 is printed on a paper surface is created. Specifically, the white pixel region in the document image is a portion where nothing is printed, and the black pixel region is a portion where black paint is applied. In this embodiment, the description will be made on the assumption that printing is performed on white paper with black ink (single color). However, the present invention is not limited to this, and printing is performed in color (multicolor). However, the present invention can be similarly applied.
[0034]
The watermark image forming unit 12 digitizes the confidential information 16 and converts it into a numerical value, performs N-ary coding (N is 2 or more), and assigns each symbol of the code word to a signal prepared in advance. A signal represents a wave having an arbitrary direction and wavelength by arranging dots in a rectangular area of an arbitrary size, and a symbol is assigned to the direction and wavelength of the wave. A watermark image is one in which these signals are arranged on the image according to a certain rule.
[0035]
The watermarked document image composition unit 13 creates a watermarked document image by superimposing the document image and the watermark image. The output device 14 is an output device such as a printer, and prints a watermarked document image on a paper medium. Therefore, the document image forming unit 11, the watermark image forming unit 12, and the watermarked document image synthesizing unit 13 may be realized as one function in the printer driver.
[0036]
The print document 20 is printed with the confidential information 16 embedded in the original document data 15 and is physically stored and managed.
[0037]
(Watermark information detection device 30)
The watermark information detection device 30 is a device that takes in a document printed on a paper medium as an image and restores embedded confidential information. As shown in FIG. 1, the watermark information detection apparatus 30 includes an input device 31 and a watermark detection unit 32.
[0038]
The input device 31 is an input device such as a scanner, and takes in the document 20 printed on paper as a multi-value gray image into a computer. The watermark detection unit 32 performs a filtering process on the input image to detect an embedded signal. The symbol is restored from the detected signal, and the embedded confidential information is extracted.
[0039]
Operations of the watermark information embedding device 10 and the watermark information detection device 30 configured as described above will be described. First, the operation of the watermark information embedding device 10 will be described with reference to FIGS.
[0040]
(Document Image Forming Unit 11)
The document data 15 is data including font information and layout information, and is created by word processing software or the like. Based on the document data 15, the document image forming unit 11 creates an image of a document printed on paper for each page. This document image is a black and white binary image. On the image, white pixels (pixels having a value of 1) are backgrounds, and black pixels (pixels having a value of 0) are character regions (regions to which ink is applied). It shall be.
[0041]
(Watermark image forming unit 12)
The confidential information 16 is various data such as characters, sounds, and images, and the watermark image forming unit creates a watermark image to be superimposed as the background of the document image from this information.
[0042]
FIG. 2 is a flowchart showing a process flow of the watermark image forming unit 12.
First, the confidential information 16 is converted into an N-element code (step S101). N is arbitrary, but in the present embodiment, N = 2 for ease of explanation. Accordingly, the code generated in step S101 is a binary code, and is expressed by a bit string of 0 and 1. In this step S101, the data may be encoded as it is, or the encrypted data may be encoded.
[0043]
Next, a watermark signal is assigned to each symbol of the codeword (step S102). The watermark signal represents a wave having an arbitrary wavelength and direction by an arrangement of dots (black pixels). The watermark signal will be further described later.
[0044]
Further, a signal unit corresponding to the bit string of the encoded data is arranged on the watermark image (step S103).
[0045]
The watermark signal assigned to each symbol of the code word in step S102 will be described. FIG. 3 is an explanatory diagram showing an example of a watermark signal.
[0046]
Let the width and height of the watermark signal be Sw and Sh, respectively. Sw and Sh may be different, but in the present embodiment, Sw = Sh is set for ease of explanation. The unit of length is the number of pixels. In the example of FIG. 3, Sw = Sh = 12. The size when these signals are printed on the paper surface depends on the resolution of the watermark image. For example, the watermark image is an image of 600 dpi (dot per inch: unit of resolution, number of dots per inch). 3, the width and height of the watermark signal in FIG. 3 is 12/600 = 0.02 (inch) on the printed document.
[0047]
Hereinafter, a rectangle having a width and a height of Sw and Sh is referred to as a “signal unit” as one signal unit. In FIG. 3 (1), the distance between dots is dense in the direction of arctan (3) (arctan is an inverse function of tan) with respect to the horizontal axis, and the wave propagation direction is arctan (−1/3). . Hereinafter, this signal unit is referred to as unit A. In FIG. 3B, the distance between dots is dense in the direction of arctan (−3) with respect to the horizontal axis, and the wave propagation direction is arctan (1/3). Hereinafter, this signal unit is referred to as unit B.
[0048]
FIG. 4 is a cross-sectional view of the change in the pixel value in FIG. 3A as viewed from the direction of arctan (1/3). In FIG. 4, the portion where the dots are arranged is the antinode of the minimum value of the wave (the point where the amplitude is maximum), and the portion where the dots are not arranged is the antinode of the maximum value of the wave.
[0049]
In addition, since there are two regions in each unit where dots are densely arranged, the frequency per unit is 2 in this example. Since the wave propagation direction is perpendicular to the direction in which the dots are densely arranged, the wave of unit A is arctan (-1/3) with respect to the horizontal direction, and the wave of unit B is arctan (1/3). Become. Note that a × b = −1 when the direction of arctan (a) and the direction of actan (b) are perpendicular.
[0050]
In the present embodiment, symbol 0 is assigned to the watermark signal expressed by unit A, and symbol 1 is assigned to the watermark signal expressed by unit B. These are called symbol units.
[0051]
In addition to the watermark signals shown in FIGS. 3 (1) and (2), for example, dot arrangements as shown in FIGS. 5 (3) to (5) are conceivable. In FIG. 5 (3), the distance between dots is dense in the direction of arctan (1/3) with respect to the horizontal axis, and the wave propagation direction is arctan (-3). Hereinafter, this signal unit is referred to as unit C.
In FIG. 5 (4), the distance between dots is dense in the direction of arctan (−1/3) with respect to the horizontal axis, and the wave propagation direction is arctan (3). Hereinafter, this signal unit is referred to as unit D. In FIG. 5 (5), the distance between dots is dense in the direction of arctan (1) with respect to the horizontal axis, and the wave propagation direction is arctan (−1). In FIG. 5 (5), it can be considered that the distance between the dots is dense in the direction of arctan (−1) with respect to the horizontal axis, and the wave propagation direction is arctan (1). Hereinafter, this signal unit is referred to as unit E.
[0052]
In this way, in addition to the combinations assigned previously, there can be a plurality of combinations of units to which symbols 0 and 1 are assigned. Therefore, it is possible to keep secret which watermark signal is assigned to which symbol. It is also possible to prevent a person (illegal person) from easily deciphering the embedded signal.
[0053]
Furthermore, when the confidential information is encoded with a quaternary code in step S102 shown in FIG. 2, for example, code A symbol 0 is assigned to unit A, symbol 1 is assigned to unit B, and symbol 2 is assigned to unit C. , It is also possible to assign symbol 3 to unit D
[0054]
In the example of the watermark signal shown in FIGS. 3 and 5, since the number of dots in one unit is all equal, by arranging these units without gaps, the apparent density of the watermark image becomes uniform. . Therefore, it appears that a gray image having a single density is embedded as a background on the printed paper.
[0055]
In order to produce such an effect, for example, unit E is defined as a background unit (signal unit to which no symbol is assigned), and this is arranged without any gap as a background of the watermark image, and symbol unit (unit A, unit B) Is embedded in the watermark image, the background unit (unit E) and the symbol unit (unit A, unit B) at the position to be embedded are exchanged.
[0056]
FIG. 6 (1) is an explanatory diagram showing a case where the unit E is defined as a background unit and arranged as a background of a watermark image without gaps. 6 (2) shows an example in which the unit A is embedded in the background image of FIG. 6 (1), and FIG. 6 (3) shows an example in which the unit B is embedded in the background image of FIG. 6 (1). Show. In the present embodiment, a method of using a background unit as the background of a watermark image will be described. However, a watermark image may be generated by arranging only symbol units.
[0057]
Next, a method for embedding one symbol of a code word in a watermark image will be described with reference to FIG.
[0058]
FIG. 7 is an explanatory diagram illustrating an example of a symbol embedding method in a watermark image. Here, a case where a bit string “0101” is embedded will be described as an example.
[0059]
As shown in FIGS. 7A and 7B, the same symbol unit is repeatedly embedded. This is to prevent the character unit in the document from being detected when a signal is detected when it is superimposed on the embedded symbol unit. The symbol unit repetition number and arrangement pattern (hereinafter referred to as a unit pattern) are used. Is optional.
[0060]
That is, as an example of the unit pattern, the repetition number is set to 4 (four symbol units exist in one unit pattern) as shown in FIG. 7 (1), or the repetition number is set to 2 as shown in FIG. 7 (2). (There are two symbol units in one unit pattern), or the number of repetitions may be one (only one symbol unit exists in one unit pattern).
[0061]
7 (1) and 7 (2), one symbol is given to one symbol unit. However, even if symbols are given to the symbol unit arrangement pattern as shown in FIG. 7 (3). good.
[0062]
The number of bits of information that can be embedded in a watermark image for one page depends on the size of the signal unit, the size of the unit pattern, and the size of the document image. The number of signals embedded in the horizontal and vertical directions of the document image may be detected as a known signal, or may be calculated backward from the size of the image input from the input device and the size of the signal unit.
[0063]
If Pw unit patterns in the horizontal direction and Ph units in the vertical direction can be embedded in a watermark image for one page, unit patterns at arbitrary positions in the image are represented by U (x, y), x = 1 to Pw, y = 1 to Ph, and U (x, y) is referred to as a “unit pattern matrix”. In addition, the number of bits that can be embedded in one page is referred to as the “number of embedded bits”. The number of embedded bits is Pw × Ph.
[0064]
FIG. 8 is a flowchart showing a method for embedding confidential information in a watermark image.
Here, a case where the same information is repeatedly embedded in one (one page) watermark image will be described. By repeatedly embedding the same information, it is possible to extract the embedded information even if the embedded information disappears due to, for example, the entire unit pattern being filled when the watermark image and the document image are overlaid. Because.
[0065]
First, the confidential information 16 is converted into an N-element code (step S201). This is the same as step S101 in FIG. Hereinafter, the encoded data is referred to as a data code, and the data code expressed by a combination of unit patterns is referred to as a data code unit Du.
[0066]
Next, based on the code length of the data code (here, the number of bits) and the number of embedded bits, how many times the data code unit can be embedded in one image is calculated (step S202). In the present embodiment, it is assumed that the code length data of the data code is inserted into the first row of the unit pattern matrix. The code length of the data code may be fixed, and the code length data may not be embedded in the watermark image.
[0067]
The number Dn of embedding the data code unit is calculated by the following equation with the data code length as Cn.
[0068]
[Expression 1]

[0069]
Here, assuming that the remainder is Rn (Rn = Cn− (Pw × (Ph−1))), the unit pattern matrix is embedded with a data pattern unit of Dn times and a unit pattern corresponding to the first Rn bits of the data code. become. However, the Rn bit of the remainder part does not necessarily have to be embedded.
[0070]
In the description of FIG. 9, the size of the unit pattern matrix is 9 × 11 (11 rows and 9 columns), and the data code length is 12 (numbers 0 to 11 in the figure indicate the codewords of the data code). And
[0071]
Next, the code length data is embedded in the first row of the unit pattern matrix (step S203). In the example of FIG. 9, the code length is expressed by 9-bit data and has been embedded only once. However, if the unit pattern matrix width Pw is sufficiently large, It can be embedded repeatedly.
[0072]
Further, the data code unit is repeatedly embedded in the second and subsequent rows of the unit pattern matrix (step S204). As shown in FIG. 9, the MSB (most significant bit) or LSB (least significant bit) of the data code is sequentially embedded in the row direction. The example of FIG. 9 shows an example in which the data code unit is embedded seven times and the first 6 bits of the data code are embedded.
[0073]
The data embedding method may be embedded so as to be continuous in the row direction as shown in FIG. 9, or may be embedded so as to be continuous in the column direction.
[0074]
The watermark image in the watermark image forming unit 12 has been described above. Next, the watermarked document image composition unit 13 of the watermark information embedding device 10 will be described.
[0075]
(Watermarked document image composition unit 13)
The watermarked document image composition unit 13 superimposes the document image created by the document image forming unit 11 and the watermark image created by the watermark image forming unit. The value of each pixel of the watermarked document image is calculated by a logical product operation (AND) of the corresponding pixel values of the document image and the watermark image. That is, if either the document image or the watermark image is 0 (black), the pixel value of the watermarked document image is 0 (black), otherwise 1 (white).
[0076]
FIG. 10 is an explanatory diagram showing an example of a watermarked document image. FIG. 11 is an explanatory diagram showing a part of FIG. 10 in an enlarged manner. Here, the unit pattern shown in FIG. 7A is used. The watermarked document image is output by the output device 14.
[0077]
The operation of the watermark information embedding device 10 has been described above.
Next, the operation of the watermark information detection apparatus 30 will be described with reference to FIGS. 1 and 12 to 20.
[0078]
(Watermark detection unit 32)
FIG. 12 is a flowchart showing the process flow of the watermark detection unit 32.
First, a watermarked document image is input to a memory of a computer or the like by an input device 31 such as a scanner (step S301). This image is referred to as an input image. The input image is a multi-valued image, and will be described below as a gray image with 256 gradations. Further, the resolution of the input image (resolution when read by the input device 31) may be different from the watermarked document image created by the watermark information embedding device 10, but here the image created by the watermark information embedding device 10 is used. It is assumed that the resolution is the same. Also, it is assumed that the input image has been corrected for rotation, expansion and contraction.
[0079]
Next, the number of unit patterns embedded is calculated from the size of the input image and the size of the signal unit (step S302). For example, assuming that the size of the input image is W (width) × H (height), the size of the signal unit is Sw × Sh, and the unit pattern is composed of Uw × Uh units. The number of unit patterns embedded in (N = Pw × Ph) is calculated as follows.
[0080]
[Expression 2]

[0081]
However, when the resolution is different between the watermark information embedding device 10 and the watermark information detection device 30, the above calculation is performed after normalizing the size of the signal unit in the input image based on the ratio of the resolutions.
[0082]
Next, a unit pattern separation position is set for the input image based on the number of unit patterns calculated in step S302 (step S303). FIG. 13 shows an example of the input image (FIG. 13 (1)) and the input image (FIG. 13 (2)) after setting the unit pattern separation positions.
[0083]
Next, a symbol unit is detected for each unit pattern delimiter, and a unit pattern matrix is restored (step S304). Details of signal detection will be described below.
[0084]
FIG. 14 is an explanatory diagram showing an example of an area corresponding to the unit A shown in FIG. 3A in the input image. In FIG. 3, the signal unit is a binary image, but here it is a multi-valued image. When a binary image is printed, since the density changes continuously due to ink bleeding or the like, the periphery of the dot becomes an intermediate color between white and black as shown in FIG. Therefore, a cross-sectional view of FIG. 14 viewed from a direction parallel to the wave propagation direction is as shown in FIG. 4 is a rectangular wave, whereas FIG. 15 is a smooth wave.
[0085]
Also, in reality, many noise components are added to the input image due to factors such as local changes in paper thickness, dirt on the printed document, and instability of the output device and image input device. However, the case where there is no noise component will be described here. However, if the method described here is used, stable signal detection can be performed even for an image to which a noise component is added.
[0086]
In the following, in order to detect the signal unit from the input image, a two-dimensional wavelet filter that can simultaneously define the frequency and direction of the wave and the influence range is used. In the following, an example using a Gabor filter, which is one of the two-dimensional wavelet filters, is shown. However, if it is a filter having the same properties as the Gabor filter, it is not necessarily a Gabor filter, and moreover, the same dot as the signal unit. A method of defining a template having a pattern and performing pattern matching may be used.
[0087]
The Gabor filter G (x, y), x = 0 to gw−1, and y = 0 to gh−1 are shown below. gw and gh are filter sizes, which are the same size as the signal unit embedded by the watermark information embedding apparatus 10 here.
[0088]
[Equation 3]

[0089]
For signal detection, the same number of Gabor filters as the frequency of the symbol unit embedded in the watermark image, the frequency, the direction of the wave, and the size are prepared as many as the type of the embedded signal unit. Here, the Gabor filters corresponding to the units A and B in FIG.
[0090]
The filter output value at an arbitrary position in the input image is calculated by convolution between the filter and the image. In the case of a Gabor filter, there are a real number filter and an imaginary number filter (an imaginary number filter is a filter whose phase is shifted by a half wavelength from the real number filter), and the mean square value thereof is used as a filter output value. For example, if the convolution between the real filter and the image of the filter A is Rc and the convolution between the imaginary filter is Ic, the output value F (A) is calculated by the following equation.
[0091]
[Expression 4]

[0092]
FIG. 16 is an explanatory diagram for explaining a method for determining whether the symbol unit embedded in the unit pattern U (x, y) delimited in step S303 is the unit A or the unit B.
[0093]
The symbol determination step for the unit pattern U (x, y) is performed as follows.
(1) While moving the position of the filter A, the maximum value of the result of calculating F (A) for all the positions in the unit pattern U (x, y) is the value of the filter A for the unit pattern U (x, y). The output value is set to Fu (A, x, y).
(2) The output value of the filter B for the unit pattern U (x, y) is calculated in the same manner as (1), and this is assumed to be Fu (B, x, y).
(3) Fu (A, x, y) and Fu (B, x, y) are compared, and if Fu (A, x, y) ≧ Fu (B, x, y), unit pattern U (x, y) The symbol unit embedded in y) is determined to be unit A, and if Fu (A, x, y) <Fu (B, x, y), it is embedded in the unit pattern U (x, y). It is determined that the existing symbol unit is unit B.
[0094]
In (1) and (2), the step width for moving the filter is arbitrary, and only the output value at a representative position on the unit pattern may be calculated. Further, when the absolute value of the difference between Fu (A, x, y) and Fu (B, x, y) is equal to or smaller than a predetermined threshold in (3), the determination may be impossible.
[0095]
In (1), when the convolution is calculated while shifting the filter, if the maximum value of F (A) exceeds a predetermined threshold value, the symbol unit immediately embedded in U (x, y) May be determined to be unit A and the process may be stopped. Similarly, in (2), if the maximum value of F (B) exceeds a predetermined threshold value, the symbol unit embedded in U (x, y) is immediately determined to be unit B. good.
[0096]
The details of the signal detection (step S304) have been described above. Returning to the flowchart of FIG. 12 again, the following step S305 will be described. In step S305, the symbols of the unit pattern matrix are concatenated to reconstruct the data code, and the original information is restored.
[0097]
FIG. 17 is an explanatory diagram showing an example of information restoration. The steps of information restoration are as follows.
(1) A symbol embedded in each unit pattern is detected ((1) in FIG. 17).
(2) The data code is restored by concatenating the symbols ((2) in FIG. 17).
(3) The embedded data is extracted by decoding the data code ((3) in FIG. 17).
[0098]
18 to 20 are explanatory diagrams illustrating an example of a data code restoration method. The restoration method is basically the reverse process of FIG.
[0099]
First, the code length data portion is extracted from the first row of the unit pattern matrix to obtain the code length of the embedded data code (step S401).
[0100]
Next, based on the size of the unit pattern matrix and the code length of the data code obtained in S401, the number Dn of data code units embedded and the remainder Rn are calculated (step S402).
[0101]
Next, data code units are extracted from the second and subsequent rows of the unit pattern matrix by the reverse method of step S203 (step S403). In the example of FIG. 19, 12 pattern units are sequentially decomposed from U (1, 2) (2 rows and 1 column) (U (1, 2) to U (3, 3), U (4, 3) to U (6, 4), ...). Since Dn = 7 and Rn = 6, 12 pattern units (data code units) are extracted 7 times, and 6 unit patterns (corresponding to the upper 6 data code units) (U (4 , 11) to U (9, 11)) are taken out.
[0102]
Next, the embedded data code is reconstructed by performing bit certainty calculation on the data code unit extracted in step S403 (step S404). Hereinafter, the bit certainty calculation will be described.
[0103]
As shown in FIG. 20, the first outer code unit extracted from the second row and first column of the unit pattern matrix is defined as Du (1,1) to Du (12,1), and Du (1,2) to Du in sequence. (12, 2),. Further, the surplus portion is assumed to be Du (1, 8) to Du (6, 8). The bit certainty calculation is to determine the value of each symbol of the data code by, for example, taking a majority vote for each element of each data code unit. As a result, even if signal detection cannot be performed correctly from any unit in any data encoding unit (such as a bit reversal error) due to overlap with the character area or dirt on the paper, the data encoding will eventually be performed correctly. Can be restored.
[0104]
Specifically, for example, the first bit of the data code is more often when the signal detection result of Du (1, 1), Du (1, 2),..., Du (1, 8) is 1. Is determined to be 1 and when there are more 0s, it is determined to be 0. Similarly, the second bit of the data code is determined by majority decision based on the signal detection result of Du (2, 1), Du (2, 2),..., Du (2, 8), and the 12th bit of the data code is Du (12,1), Du (12,2),..., Du (12,7) (Du (12,8) is not present, so it is up to Du (12,7)). Judgment by majority vote.
[0105]
The bit certainty calculation can also be performed by adding the output values of the signal detection filter of FIG. This is because, for example, a symbol of 0 is assigned to unit A in FIG. 3 (1), a symbol of 1 is assigned to unit B in FIG. 3 (2), and filter A for Du (m, n) is used. When the maximum value of the output value is Df (A, m, n) and the maximum value of the output value by the filter B for Du (m, n) is Df (B, m, n), the M bit of the data code is
[0106]
[Equation 5]

[0107]
In the case of, it is determined as 1, and in other cases it is determined as 0. However, when N <Rn, the addition of Df is n = 1 to Rn + 1.
[0108]
Although the case where data codes are repeatedly embedded has been described here, a method that does not repeat data code units can be realized by using an error correction code or the like when encoding data.
[0109]
As described above in detail, according to the present embodiment, the following excellent effects are obtained.
(1-1) Since the embedded information is expressed by the difference in dot arrangement, the original document font, character spacing, and line spacing are not changed.
(1-2) Since the dot pattern to which symbols are assigned and the dot pattern to which symbols are not assigned have the same density (the number of dots in a certain interval), the human eye has a uniform density on the background of the document. It appears to be shaded and the presence of information is inconspicuous.
(1-3) Keeping the dot pattern to which symbols are assigned and the dot pattern to which symbols are not assigned in secret makes it difficult to decipher the embedded information.
(1-4) A pattern representing information is a collection of fine dots that are embedded as a background of a document, so that even if an embedding algorithm is disclosed, it is difficult to falsify embedded information in a printed document. .
(1-5) Since the embedded signal is detected based on the difference in the direction of the wave (change in shading) (because detailed detection is not performed in units of one pixel), the printed document is stable even if there is some dirt, etc. Information detection can be performed.
(1-6) Since the same information is repeatedly embedded and information is restored using all of the repeatedly embedded information at the time of detection, the signal portion is hidden by large font characters or the paper is dirty. Even if partial information loss occurs, the embedded information can be taken out stably.
[0110]
(Second Embodiment)
In the document output unit of the first embodiment, data to be inserted into the printed material as a watermark is converted into a code word (this code word may be encrypted) using an error correction code or the like. For example, when converted to a binary code, data is represented by a bit string of 0 and 1. Each symbol (symbol 0, symbol 1) corresponds to a dot pattern representing it. This is embedded as the background of the document image, but when the number of bits of the code word (number of elements of the dot pattern array) is smaller than the maximum number of signals that can be embedded on one sheet of paper, the dot pattern array It will be embedded repeatedly.
[0111]
Since the watermark signal is embedded as the background of the document image, the dot pattern cannot be detected at the time of signal detection from a place where part or all of the dot pattern representing the signal overlaps the character area of the document image. As described in the first embodiment, when a signal is detected, whether each bit of the code word is 0 or 1 is determined by a dot pattern detected from a position corresponding to each bit. It depends on the method such as majority vote. Accordingly, when the position where the dot pattern cannot be detected (the area overlapping the character area) is concentrated on a certain bit of the code word, the bit cannot be determined as shown in FIG. Become. Even if an error correction code is used for a code word, information cannot be restored if an undecidable bit exceeds the error correction capability.
[0112]
FIG. 22 is an explanatory diagram showing the configuration of the watermark information embedding device and the watermark information detection device according to this embodiment. The difference from the first embodiment is that an embedding pattern determination unit 42 is added to the watermark information embedding device 40. The embedding pattern determination unit 42 is a part that determines in what order the watermark signals are arranged on the paper.
[0113]
Since the other components are the same as those in the first embodiment, the redundant description will be omitted, and the operation of the present embodiment will be described focusing on the additions to the first embodiment. To do.
[0114]
FIG. 23 shows a processing flow of the embedding pattern determination unit 42 and the watermark image forming unit 12. A region surrounded by a broken line (steps S501 to S505) is a portion corresponding to the embedding pattern determination unit 42.
[0115]
The outline of the embedding pattern determination unit 42 is as follows. When Np types of signal embedding patterns are prepared in advance and embedded in accordance with the respective embedding patterns before actual signal placement, the extent to which the character overlaps is determined. It is determined whether or not to concentrate on a specific bit at a rate of, and an embedding pattern is selected so that a portion overlapping with a character is most widely dispersed in all bits of the code word.
[0116]
When the code length is L and the number of signals (dot patterns) that can be embedded in the document is N, the number of effective embeddings of each bit of the codeword (embedding without overlapping with characters) The average value V of the number of times that can be performed is N / L. Therefore, ideally, the embedding pattern that can be arranged in an area where all the bits of the code word do not overlap the character V times is the best pattern. As described above, if the number of effective embeddings is biased, bits that could be embedded only a small number of times will not be read due to signal detection errors during signal detection, dirt on the paper after printing, new writing on the paper, etc. It is likely to be impossible. If the bias is large, there are bits that are not affected at all by dirt and writing, but many bits that cannot be read out appear. Therefore, it is only necessary to select a pattern in which all the bits of the code word are close to the average value of the effective embedding frequency, that is, the variance (or standard deviation) of the effective embedding frequency of each bit is smallest.
[0117]
In step S501, data to be inserted into the printed material as a watermark is converted into a code word (this code word may be encrypted) using an error correction code or the like. For example, when converted to a binary code, data is represented by a bit string of 0 and 1. Each symbol (symbol 0, symbol 1) corresponds to a dot pattern representing it.
[0118]
In steps S502 to S504, for the Np types of embedding patterns (step S502), the number of cases where each bit of the code word to be embedded overlaps the character area is counted for each bit (step S503).
[0119]
FIG. 24 shows examples of four types of embedding patterns (Np = 4). In FIG. 24A, signals are continuously embedded in the horizontal direction, (b) is continuously embedded in the vertical direction, and (c) and (d) are embedded continuously in the oblique direction. There are countless other possibilities, such as embedding in a spiral shape. Further, as shown in FIG. 25, the document image may be divided into several blocks, and different embedding patterns may be used for each block.
[0120]
FIG. 26 shows an example of codewords and an example in which dot patterns are arranged based on them. This dot pattern is repeatedly arranged as the background of the document image.
[0121]
FIG. 27A shows an example in which the code words of FIG. 26 are arranged according to the embedding pattern of FIG. Here, the m-th dot pattern in FIG. 26 is embedded 12 times. FIG. 27B shows a print pattern superimposed on the dot pattern foreground. As shown in the figure, the m-th bit of the code word does not overlap the character in the first arrangement, but overlaps the character in the second arrangement. For the embedding pattern P in FIG. 24, the number when the m-th bit of the code word can be arranged without overlapping the character is represented as N (P, m). The total number T (P) of effective embedding times of each bit of the code word is obtained by adding N (P, m) from m = 1 to L, and the average effective embedding number E (P) is T (P) / L. It becomes. From this, the distribution of the effective embedding frequency is
V (P) = E {E (P) -N (P, m)) ² }
It can be calculated with
[0122]
V (P) is calculated for all the embedding patterns, and the embedding pattern having the smallest V (P) is determined to be the embedding pattern for the document image in step S505.
[0123]
In step S506, a watermark signal is arranged according to the embedding pattern determined in step S505, and is superimposed on the document image. At this time, as shown in FIG. 28, information on which embedding pattern is embedded in the header information area is simultaneously embedded. It is assumed that information is arranged in the header information area with a fixed embedding pattern regardless of the document.
[0124]
Next, the watermark detection unit 32 will be described.
FIG. 29 shows the processing flow of the watermark detection unit 32.
In step S601, the signal of the portion where the header information is embedded in the input image is detected, and the header information is restored.
In step S602, a signal embedding pattern is extracted from the header information restored in step S601.
In step S603, the watermark signal is read according to the embedding pattern extracted in step S602.
In step S604, a majority vote is taken for each bit of the code word, the value of each bit is determined, and the code is combined to extract information.
[0125]
In step S506 of FIG. 23, the embedded pattern is recorded in the header information embedded area. However, if the embedded pattern is publicly available and is limited to several types, all patterns are detected at the time of detection without recording the embedded pattern. It is also possible to extract information that has been decoded correctly (no decoding error has occurred or meaningful information has been extracted) as watermark information.
[0126]
Even when the embedded pattern is kept secret, it is not necessary to record the embedded pattern. This is more secure by making the embedding pattern more complex and not revealing what the pattern is, and keeping the embedded signal pattern secret when the watermark information is secret information. The purpose is to keep sex. When used for such a purpose, it is not always necessary to perform the processing of steps S502 to S505.
[0127]
In addition, a supplementary explanation about block division (FIG. 25) of an embedding pattern is that when a document image is divided into several blocks as one of signal embedding patterns and different embedding patterns are applied to the respective blocks, As a division method, a method as described below may be used in addition to division with a fixed-size block as shown in FIG.
[0128]
As a premise, for example, when a horizontally embedded embedding pattern (FIG. 24 (a)) is adopted in a horizontally written document, an area overlapping with a character continues to some extent, so depending on the code length, as shown in FIG. Situations where it cannot be determined are likely to occur. The same applies to the case where an embedded pattern continuous in the vertical direction (FIG. 24B) is adopted for a vertically written document.
[0129]
FIG. 30 shows an example of a document in which vertical writing, horizontal writing, and drawings are mixed. In such a document, rather than applying one embedding pattern to the entire document as shown in FIG. 24, dividing each block into several blocks and applying a different embedding pattern to each block makes it possible to The variance of the effective number of embeddings is reduced. The division method may be divided into blocks having a fixed size as described above. However, the effect of reducing the variance value is greater if dynamic division is performed by analyzing the characteristics of the document image.
[0130]
FIG. 31A shows an example in which the document of FIG. 30 is divided according to features such as vertical writing and horizontal writing. This division may use a method such as “OCR table analysis (Japanese Patent Laid-Open No. 04-033079)” used as a pre-processing of OCR, or a character area of a certain size by reducing the document image of FIG. It is also possible to extract a lump and divide the lump as one block. In the example of FIG. 31A, areas 1, 2, and 4 are horizontally written character string areas, area 5 is a vertically written character string area, and area 3 is the area shown in the figure.
[0131]
For each of these areas, steps S502 to S505 in FIG. 23 are performed to determine an embedding pattern for each area. FIG. 31B shows an example in which an optimum embedding pattern is applied to each region.
[0132]
Information on how the document image is divided and information on the embedding pattern for each area may be recorded in the header information area of FIG. 28, or may be secret information.
[0133]
Further, since it is difficult to embed a signal in the figure area in the document image, when dividing the area in FIG. 31A, the signal is not embedded in the area determined as the figure area, and the signal is not embedded in the figure area. Information that it was not embedded may be recorded in the header information area. By not embedding the signal in the figure area, it is possible to omit the process of detecting the signal from the area where the signal is hardly embedded at the time of watermark detection, so that the processing time can be shortened and signal detection errors can be prevented. .
[0134]
As described above in detail, according to the present embodiment, the following excellent effects are obtained.
(2-1) When embedding information in a method that does not give a sense of incongruity with a minute dot pattern on a paper background, a signal that can be detected by the overlap of the foreground character area and the dot pattern is a code representing information. Since it does not concentrate on some bits of a word, it is possible to reliably extract information even when a signal error is detected, deterioration such as paper contamination, or overwriting on paper.
(2-2) By keeping the signal embedding pattern secret, only those who know the embedding pattern can retrieve the information, and the risk of the embedded information being stolen by a malicious third party Sex is reduced.
(2-3) The effect of (2-1) can be further enhanced by dividing the document image into regions for each component feature and applying an optimum signal embedding pattern for each divided region.
(2-4) The document image is divided into regions for each component feature, and the signal is not embedded in parts such as figure regions where signals cannot be embedded, thereby shortening the processing time during watermark detection and detection. Errors can be reduced.
[0135]
The preferred embodiments of the watermark information embedding device and the watermark information detection device according to the present invention have been described above with reference to the accompanying drawings, but the present invention is not limited to such examples. It will be obvious to those skilled in the art that various changes or modifications can be conceived within the scope of the technical idea described in the claims, and these are naturally within the technical scope of the present invention. It is understood that it belongs.
[0136]
【The invention's effect】
As described above, according to the present invention, the following excellent effects are obtained.
(1-1) Since the embedded information is expressed by the difference in dot arrangement, the original document font, character spacing, and line spacing are not changed.
(1-2) Since the dot pattern to which symbols are assigned and the dot pattern to which symbols are not assigned have the same density (the number of dots in a certain interval), the human eye has a uniform density on the background of the document. It appears to be shaded and the presence of information is inconspicuous.
(1-3) Keeping the dot pattern to which symbols are assigned and the dot pattern to which symbols are not assigned in secret makes it difficult to decipher the embedded information.
(1-4) A pattern representing information is a collection of fine dots that are embedded as a background of a document, so that even if an embedding algorithm is disclosed, it is difficult to falsify embedded information in a printed document. .
(1-5) Since the embedded signal is detected based on the difference in the direction of the wave (change in shading) (because detailed detection is not performed in units of one pixel), the printed document is stable even if there is some dirt, etc. Information detection can be performed.
(1-6) Since the same information is repeatedly embedded and information is restored using all of the repeatedly embedded information at the time of detection, the signal portion is hidden by large font characters or the paper is dirty. Even if partial information loss occurs, the embedded information can be taken out stably.
[0137]
In addition, according to the present invention, there are the following excellent effects.
(2-1) In the case where information is embedded in a method that does not give a sense of incongruity with a minute dot pattern on a paper background, a signal that can be detected by the overlap of the foreground character area and the dot pattern is a code representing information. Since it does not concentrate on some bits of a word, it is possible to reliably extract information even when a signal is detected, a detection error is deteriorated, a paper stain is deteriorated, or the paper is overwritten.
(2-2) By keeping the signal embedding pattern secret, only those who know the embedding pattern can retrieve the information, and the risk of the embedded information being stolen by a malicious third party Sex is reduced.
(2-3) The effect of (2-1) can be further enhanced by dividing the document image into regions for each component feature and applying an optimum signal embedding pattern for each divided region.
(2-4) The document image is divided into regions for each component feature, and the signal is not embedded in parts such as figure regions where signals cannot be embedded, thereby shortening the processing time during watermark detection and detection. Errors can be reduced.
[Brief description of the drawings]
FIG. 1 is an explanatory diagram illustrating a configuration of a watermark information embedding device and a watermark information detection device according to a first embodiment;
FIG. 2 is a flowchart showing a processing flow of a watermark image forming unit 12;
FIG. 3 is an explanatory diagram showing an example of a watermark signal, where (1) shows unit A and (2) shows unit B;
4 is a cross-sectional view of the change in pixel value in FIG. 3A as viewed from the direction of arctan (1/3).
FIGS. 5A and 5B are explanatory diagrams illustrating an example of a watermark signal. FIG. 5 illustrates a unit C, FIG. 5 illustrates a unit D, and FIG. 5 illustrates a unit E. FIG.
FIG. 6 is an explanatory diagram of a background image. (1) shows a case where unit E is defined as a background unit and this is used as the background of a watermark image arranged without gaps, and (2) shows the background of (1). An example in which the unit A is embedded in the image is shown, and (3) shows an example in which the unit B is embedded in the background image of (1).
FIG. 7 is an explanatory diagram illustrating an example of a symbol embedding method in a watermark image.
FIG. 8 is a flowchart showing a method of embedding confidential information in a watermark image.
FIG. 9 is a flowchart showing a process flow of the watermark detection unit 32;
FIG. 10 is an explanatory diagram illustrating an example of a watermarked document image.
FIG. 11 is an explanatory view showing a part of FIG. 10 in an enlarged manner.
12 is a flowchart showing a process flow of the watermark detection unit 32. FIG.
FIG. 13 shows an example of an input image (FIG. 13 (1)) and an input image (FIG. 13 (2)) after setting a unit pattern separation position.
FIG. 14 is an explanatory diagram showing an example of a region corresponding to unit A in an input image.
15 is a cross-sectional view of FIG. 14 viewed from a direction parallel to the wave propagation direction.
FIG. 16 is an explanatory diagram for explaining a method for determining whether a symbol unit embedded in a unit pattern U (x, y) is a unit A or a unit B;
FIG. 17 is an explanatory diagram showing an example of information restoration
FIG. 18 is an explanatory diagram showing an example of a data code restoration method;
FIG. 19 is an explanatory diagram illustrating an example of a data code restoration method;
FIG. 20 is an explanatory diagram illustrating an example of a data code restoration method;
FIG. 21 is an explanatory diagram showing a case where it becomes impossible to detect a dot pattern.
FIG. 22 is an explanatory diagram illustrating a configuration of a watermark information embedding device and a watermark information detection device according to a second embodiment;
FIG. 23 is a flowchart showing a processing flow of the embedding pattern determination unit.
FIGS. 24A and 24B are explanatory diagrams showing examples of embedding patterns, where FIG. 24A is a horizontal direction, FIG. 24B is a vertical direction, FIG. 24C is an oblique direction from upper right to lower left, and FIG. The diagonal direction is shown.
FIG. 25 is an explanatory diagram showing a case where a document image is divided into regions.
FIG. 26 is an explanatory diagram illustrating an example of codewords and an example in which dot patterns are arranged based on the codewords.
27 is an explanatory diagram showing an example in which the code words of FIG. 6 are arranged according to the embedding pattern of FIG.
FIG. 28 is an explanatory diagram showing a header information area and a watermark information area.
29 is a flowchart showing a process flow of the watermark detection unit 32. FIG.
FIG. 30 is an explanatory diagram illustrating an example of a document in which vertical writing, horizontal writing, drawings, and the like are mixed.
FIG. 31 is an explanatory diagram illustrating an example in which the document in FIG. 30 is divided according to features such as vertical writing, horizontal writing, and drawing.
[Explanation of symbols]
10 Watermark information embedding device
11 Document image forming section
12 Watermark image forming part
13 watermarked document image composition unit
14 Output device
15 Document data
16 Confidential information
20 Printed documents
30 Watermark information detection device
31 input devices
32 Watermark detection unit
40 Watermark information embedding device
42 Embedding pattern determination unit

Claims (20)

In the watermark information embedding device,
It has a watermarked image composition unit that creates a watermarked image by overlaying dot patterns on the image,
In the dot pattern , dots are arranged at predetermined intervals in a predetermined direction, the luminance value of the dot is an amplitude, and a direction perpendicular to the predetermined direction is a propagation direction,
A watermark information embedding apparatus characterized in that a plurality of dot patterns with different dot directions and / or intervals are prepared, and confidential information representing a specific secret is given by arranging the dot patterns in combination. . 2. The watermark information embedding device according to claim 1, wherein one symbol is given to one dot pattern. 2. The watermark information embedding apparatus according to claim 1, wherein one symbol is given to a combination of a plurality of dot pattern arrangements. 4. The watermark information embedding apparatus according to claim 2 , wherein the symbols include a symbol that forms part of the confidential information and a symbol that is not related to the confidential information. And an output device for outputting the watermarked image on paper.
5. The watermark information embedding apparatus according to claim 1, wherein the confidential information is repeatedly embedded in an arbitrary range on paper output by the output device. And an output device for outputting the watermarked image on paper.
5. The watermark information embedding apparatus according to claim 1, wherein the confidential information is repeatedly embedded in a part or the entire surface of paper output by the output device. In the watermark information detection device,
A watermark detection unit for detecting the watermark image from a watermarked image created by superimposing an image and a watermark image in which a plurality of types of dot patterns are embedded;
In the dot pattern, dots are arranged at predetermined intervals in a predetermined direction, the luminance value of the dot is an amplitude, and a direction perpendicular to the predetermined direction is a propagation direction,
By preparing a plurality of dot patterns in which the direction and / or interval of the dots are changed and arranging the dot patterns in combination, confidential information indicating a specific confidentiality is given,
The watermark detection unit includes the same number of two-dimensional wavelet filters as the dot pattern having the same direction and interval as the dot pattern, and detects the watermark image by performing matching on the watermarked image. A watermark information detection device characterized by In addition, it has an input device that reads watermarked images printed on paper.
8. The watermark information detection apparatus according to claim 7, wherein the watermark detection unit detects the watermark image by performing matching on the watermarked image read by the input device. A convolution of an arbitrary region in the watermarked image and the plurality of two-dimensional wavelet filters is calculated, and the dot pattern corresponding to the two-dimensional wavelet filter maximizing the convolution is embedded in the region The watermark information detecting apparatus according to claim 7 or 8 , wherein the watermark information detecting apparatus is determined. When the convolution of an arbitrary area in the watermarked image and an arbitrary two-dimensional wavelet filter exceeds a certain threshold, the dot pattern corresponding to the two-dimensional wavelet filter is embedded in the area The watermark information detecting apparatus according to claim 7 or 8 , wherein the watermark information detecting apparatus is determined. The confidential information is repeatedly embedded in an arbitrary range or on the entire surface of the paper input to the input device,
The convolution with all signal detection filters at the position corresponding to the same sign bit in the watermarked image is added for each of the plurality of two-dimensional wavelet filters by the number of repetitions at the time of embedding, and the added value is the largest. 9. The watermark information detecting apparatus according to claim 7 , wherein the dot pattern corresponding to the two-dimensional wavelet filter is determined to be embedded. The watermark information detecting apparatus according to claim 7, 8, 9, 10, or 11 , wherein the two-dimensional wavelet filter is a Gabor filter. Characterized by comprising the plural types of dot patterns what embedding further a pattern determination unit for selecting whether to place a pattern of watermark according to any one of claims 1, 2, 3, 4 or 5 Information embedding device. The pattern determination unit
Change the embedding pattern of the dot pattern by two or more, and temporarily arrange the dot pattern;
14. The watermark according to claim 13 , wherein the embedding pattern is selected according to the number of dot patterns (number of effective dot patterns) that can be detected even if the image and the watermark image are overlapped. Information embedding device. The pattern determination unit
For two or more of each embedding pattern subjected to the temporary arrangement, wherein to calculate the effective dot pattern number of the variance of each dot pattern, variance and selects the smallest the embedding pattern, the claim 14 The watermark information embedding device described. 16. The watermark information embedding device according to claim 13 , wherein information on the embedding pattern selected by the pattern determination unit is embedded as header information in the watermarked image. The watermark information embedding apparatus according to claim 16 , wherein the information on the embedding pattern selected by the pattern determination unit is encrypted and embedded as header information in the watermarked image. The pattern determination unit
18. The watermark information embedding apparatus according to claim 13 , wherein the image is divided into regions, and the optimum embedding pattern is selected for each divided region. The pattern determination unit
19. The watermark information embedding apparatus according to claim 18 , wherein characteristic recognition processing is performed on the image, and region division is performed according to the characteristic of the image. The pattern determination unit does not place the dot pattern in a divided region where the dot pattern cannot be substantially arranged as a result of the characteristic recognition process on the image,
20. The watermark information embedding apparatus according to claim 19 , wherein the information regarding the divided areas where the dot pattern is not arranged is embedded as header information in the watermarked image.

Images