JP3718967B2 - Image feature amount extraction apparatus and method, and recording medium on which image feature amount extraction program is recorded - Google Patents

Description

ここで、Ｓ _θ （ i,j ｜ｄ）は行列Ｓ _θ （ｄ）のi行、j列要素であり、Ｎ_Gは画像の濃度レベルの数である。
【００２３】
テクスチャの周期性は濃度の周期性であるから、テクスチャの周期性はＳ_θ（i,j｜ｄ）に反映されるはずである。そこで、各方向に対してｄを３、４、…ｄ_max（距離ｄの最大値）と変化させた場合のグラフから周期性の有無を判定する。以下にこの判定方法について説明する。まず、各方向θにおいて、各距離ｄに対する慣性Ｉ{Ｓ_θ（ｄ）}を求め、各慣性Ｉ{Ｓ_θ（ｄ）}の中から最小の慣性の値Ｉner（θ）を求める。ここで、Ｉner（θ）は次式で与えられる。

【００２４】
その後、Ｉner（０）、Ｉner（４５）、Ｉner（９０）、Ｉner（１３５）の中から、最小のものと、２番目に小さいものとを選択する。選択されたそれぞれの慣性の最小値Ｉner（θ）が所定値より小さければ、対象とするテクスチャが周期性を持つと判断する。
【００２５】
例えば、図５は、４つの方向θにおいて、ある画像に対して距離ｄを変化させた場合の慣性Ｉを求めた結果を示したものである。この図の場合、方向θ＝０°では距離ｄ＝７のときに、方向θ＝４５°では距離ｄ＝７のときに、方向θ＝９０°では距離ｄ＝１３のときに、方向θ＝１３５°では距離ｄ＝６のときに、それぞれの方向での慣性の最小値Ｉner（θ）が求まる。各方向θに対する慣性Ｉの最小値Ｉner（θ）のうち、Ｉner（１３５）が最小となり、Ｉner（４５）が２番目に小さい値となる。そして、Ｉner（４５）、Ｉner（１３５）の双方とも周期性判定の閾値より低いため、この画像はθ＝４５°、１３５°の方向に対して周期性があると判断される。
【００２６】
＜画像特徴量抽出装置の制御動作＞
以下に、本システムの具体的な制御動作についてフローチャートを用いて説明する。
【００２７】
＜メインフロー＞
図６は本システムにおいて実行されるプログラムのメインルーチンを示すフローチャートである。本プログラムが起動されると、まず、以降の各処理で必要なフラグ等のイニシャライズや、初期メニュー画面の表示等を行う初期設定処理が行われる（Ｓ１）。図７に初期画面の一例を示す。初期メニュー画面２１上では、所定の処理を選択するための選択項目２３〜２５がアイコンとして表示されており、この選択項目２３〜２５の１つがユーザにより選択されることにより所定の処理が実行される。なお、本システムにおいては、ディスプレイ２等上に表示された初期メニュー画面２１等の設定画面上で、ユーザにより、キーボード３やマウス４等を介して各種処理の選択、設定値の入力等が行われる。ステップＳ１の後、初期メニュー画面２１上でユーザによるメニュー選択がなされたか否かを判定する（Ｓ２）。ステップＳ２において、「テクスチャ抽出」２３が選択されれば、指定された画像データからテクスチャ特徴量を抽出し、画像データベース５０に登録する等の処理を行うテクスチャ抽出処理（Ｓ３）へ進み、その後、ステップＳ６へ進む。ステップＳ２において、「テクスチャ比較検索」２４が選択されれば、指定された画像データと、データベース５０に登録されている画像データのテクスチャを比較し、類似する画像データを抽出する処理を行うテクスチャ比較検索処理（Ｓ４）へ進み、その後、ステップＳ６へ進む。ステップＳ２において、「その他のメニュー」２５が選択されれば、その他のメニュー処理を行い（Ｓ５）、その後、ステップＳ６へ進む。ステップＳ２において、メニュー選択されなければ、なにもせずにステップＳ６へ進む。ステップＳ６では、その他の処理を実行し、すべての処理が終わるとステップＳ２へ戻り、以後、同様の処理が繰り返される。
【００２８】
ここで、その他のメニュー処理（Ｓ５）についても、一般的な検索システムと基本的に同様であり、本願発明に直接関係しないのでここでの説明は省略し、特に、テクスチャ抽出処理（ステップＳ３）およびテクスチャ比較検索処理（ステップＳ４）について以下に詳細に説明する。
【００２９】
＜画像特徴量抽出装置におけるテクスチャ抽出処理＞
最初に、テクスチャ抽出処理（ステップＳ３）について説明する。図８にテクスチャ抽出処理のフローチャートを示す。図８に示すようにテクスチャ抽出処理では、画像データを画像データベース５０に登録する際に、画像データから濃度共起行列を計算する濃度共起行列計算処理（Ｓ３１）と、濃度共起行列計算処理の結果に基づいてテクスチャ特徴量を抽出し、この特徴量を画像データとともに画像データベース５０に登録する特徴量抽出処理（Ｓ３２）とを行う。以下に、それぞれのステップを図９及び図１０のフローチャートを参照して説明する。
【００３０】
図９は濃度共起行列計算処理（Ｓ３１）のフローチャートである。図９において、まず、データベースやファイルシステムなどの記憶媒体から画像をロードし（Ｓ３０１）、正規化する（Ｓ３０２）。ここでの正規化とは、画像処理時間の短縮とノイズの削減の目的のために、取得する画像を所定の大きさに縮小または拡大することである。本実施形態では１２０×１２０ピクセルの大きさに画像が収まるように縦横比を維持したまま縮小する。
【００３１】
次に、正規化されたデータをグレー画像に変換した後、二値化処理を行う（Ｓ３０３）。ここで、本実施形態では、従来技術で述べた二値化処理の問題を解決すべく、閾値を出現頻度の高い画素値に設定せず、その前後の値を閾値に設定するようにした。すなわち、閾値をグレー化後の画素値分布の中間値から所定値だけ前後にシフトした値に設定する。以下に具体例を用いて説明する。
【００３２】
図１１はグレー化後の画像の画素値に対する画素値分布を示したものである。この図では画素値は２５６諧調を有する。本実施形態では、図１１に示すように、画素値分布において、画素値の中間値（Ｍ）から、標準偏差（ｓ）の３分の１だけ画素値の高い方にシフトした値に閾値を設定する。このように、二値化の閾値を画素値分布の中間値から所定値だけシフトした値に設定することにより、頻度の高い画素値に閾値を設定しないため、元画像の特徴をより保持したまま二値化が可能となる。ただし、以下の場合は、上記とは異なる閾値の設定を行う。すなわち、中間値Ｍが１２８以上で、かつ、中間値での累積画素数が９０％を越える場合は、画素値が中間値未満となる画素を「０」にし、中間値以上となる画素を「１」にする。また、中間値Ｍが１２８未満で、かつ、中間値での累積画素数が９０％を越える場合は、画素値が中間値以下となる画素を「０」にし、中間値を越える画素を「１」にする。また、（中間値Ｍ＋標準偏差ｓ／３）となる値での累積画素数が１００％になる場合は、閾値は中間値Ｍに設定する。このように閾値を設定し二値化処理することにより、結果として元画像のイメージに近い二値化画像が得られる。
【００３３】
画像の正規化（Ｓ３０２）及び二値化（Ｓ３０３）が終了すると、次に、当該画像から濃度共起行列を計算する。まず、濃度共起行列を求める走査方向θに対応した方向番号ｎを初期値０に設定する（Ｓ３０４）。方向番号ｎは、０のとき０°、１のとき４５°、２とき９０°、３のとき１３５°の方向θにそれぞれ対応する。ここで、より正確な方向性を求めるならば、例えば、濃度共起行列を求める角度間隔を１０°毎に設定し、方向番号が０のとき０°、１のとき１０°、２のとき２０°というように角度を細かく分割して走査方向θに対応させ、濃度共起行列を求めてもよい。しかし、角度を細かく分割すると計算処理が多くなり、処理時間がかかるため、ここでは４５°づつに分割する。
【００３４】
次に、濃度共起行列を求める時の画素の距離ｄを１に設定する（Ｓ３０５）。第ｎ方向、距離ｄにおける濃度共起行列を求める（Ｓ３０６）。すなわち、この場合、第１方向、距離１における濃度共起行列を求めることになる。本実施形態においては、濃度共起行列を計算する際には、画像全体の画素に対して求めるのではなく、計算対象の方向と同じ方向の所定幅の帯上の画素に対して行う。帯の幅は画像の大きさに比べて処理の高速化の効果を得るために充分小さく、かつ濃度共起行列が得られるため必要な大きさであれば良い。当然、帯の幅を太くすれば計算結果の精度は上がるが計算対象の画素が増えるので処理は遅くなり、逆に帯を細くすれば処理は速くなるが計算結果の精度は下がる。
【００３５】
本実施例では正規化の画像サイズが１２０×１２０であることを考慮して、帯の幅を１０ピクセルとして説明していく。例えば、第０方向であれば、図４の（ａ）に示すように画像の中心を通り１０ピクセル幅の横方向の帯になる。第１方向であれば、図４の（ｂ）に示すように画像の中心を通り１０ピクセル幅の４５°方向の帯になる。第２方向であれば、図４の（ｃ）に示すように画像の中心を通り１０ピクセル幅の９０°方向の帯になる。第３方向であれば、図４の（ｄ）に示すように画像の中心を通り１０ピクセル幅の１３５°方向の帯になる。なお、走査する画像領域は、図４に示すように連続でなくてもよく、同一方向性を有していれば、不連続な複数の線または領域からなる領域であってもよい。例えば、離散的に位置する１ピクセルづつの１０本の線からなる領域を走査してもよい。画像全体に周期性があれば、このようにしても同様に周期性は求められるからである。
【００３６】
ステップＳ３０６で得られた濃度共起行列から式（５）を用いて慣性の値Ｉ{Ｓ_θ（ｄ）｝を求める（Ｓ３０７）。ステップＳ３０７で得られた慣性の値を配列Ｖａｌ［ｄ］に代入する（Ｓ３０８）。その後、距離ｄの値をインクリメントする（Ｓ３０９）。
【００３７】
設定された距離ｄが、所定の距離の上限ｄ_maxを越えているか否かを判定する（Ｓ３１０）。距離ｄが上限ｄ_maxを越えていればＳ３１１に進む。距離ｄがその上限ｄ_max以下であればステップＳ３０６に戻り、距離ｄがその上限ｄ_maxを越えるまで、処理（Ｓ３０６〜Ｓ３１０）を繰り返す。ここで、距離の上限ｄ_maxは、画像のサイズの半分程度の値までに設定する。
【００３８】
ステップＳ３１０では、同一方向における、各距離ｄに対する慣性Ｖａｌ[ｄ]の値の中から、その方向における慣性の最小値Ｉner[ｎ]を求める（Ｓ３１１）。
【００３９】
その後、次の方向を指定するため、方向番号ｎをインクリメントする（Ｓ３１２）。インクリメントされた方向番号ｎが３以下であるかを判定する（Ｓ３１３）。方向番号ｎが３以下であれば処理Ｓ３０５に戻り、上記処理を繰り返す（Ｓ３０５〜Ｓ３１３）。全ての方向番号ｎに対して慣性の値の計算が終了すると、処理を終了する。
【００４０】
以上のようにして、濃度共起行列を用いて所定の方向に対する画像の慣性を求めると、次にこの慣性の値に基づいて画像の特徴量を抽出する。ここで、本実施形態では、画像のテクスチャの周期性をテクスチャ特徴量として抽出する。一般に、濃度共起行列から得られた慣性の値は同じ方向で見た場合、テクスチャの周期性が顕著である場合ほど距離ｄに対する慣性の値に周期性が見られることは良く知られている。そして、周期となる距離ｄでは慣性の値が０に近づく。このため、本実施形態では、周期性を判断するための閾値を決め、この閾値と慣性の最小値とを比較し、慣性の最小値が閾値より小さいときに、慣性の最小値が与えられる距離ｄにおいて周期性があると判断する。
【００４１】
図１０は特徴量抽出処理（Ｓ３２）のフローチャートである。図１０において、各走査方向に対して慣性の最小値を格納したＩner［０］〜Ｉner［３］のうちから最小値を選択する(Ｓ３１４)。そして、この最小値が前述の周期性を判断するための閾値以下であるか否かを判断する(Ｓ３１５)。閾値を越えていればステップＳ３２２に進み、閾値以下であればステップＳ３１６へ進む。
【００４２】
ステップＳ３１６では、周期性があると判断された最小値の慣性に対する方向番号ｎと周期ｄをテクスチャ特徴量として保存する。ここで、最小の慣性の最小値を与える周期性を「第１の周期」と呼ぶ。また、ここで得られたテクスチャ特徴量の方向と周期を第１周期特徴量と呼ぶ。
【００４３】
次に、４つの走査方向に対する各慣性の最小値を格納した配列Ｉner［０］〜Ｉner［３］のうちから２番目に小さい慣性の値を選択する（Ｓ３１７）。この慣性の値が、周期性を持つと判断する閾値以下であるか否かを判断し（Ｓ３１８）、閾値を越えていればステップＳ３２２に進む。閾値以下であればステップＳ３１９へ進む。
【００４４】
ステップＳ３１９では、周期性があると判断された２番目に小さい値の慣性に対する方向番号ｎと周期ｄをテクスチャ特徴量として保存する。ここで、２番目に小さい慣性の最小値を与える周期性を「第２の周期」と呼ぶ。また、ここで得られたテクスチャの方向性と周期を第２周期特徴量と呼ぶ。
【００４５】
その後、ステップＳ３２０において、周期性フラグを「有」にセットし、処理を終了する。ここで、周期性フラグは「有」のとき対象画像に周期性があることを意味し、「無」のとき対象画像に周期性がないことを意味する。ステップＳ３２２では、周期性フラグを「無」にセットする。最後に、画像データを周期性フラグおよび特徴量（周期性があるとき）とともに画像データベース５０に登録し（Ｓ３２１）、処理を終了する。
【００４６】
以上の処理において、周期性フラグにより対象画像が周期性を有するか否かの情報が得られ、さらに周期性がある場合は、「第１周期」の方向と周期及び「第２周期」の方向と周期がテクスチャ特徴量として得られる。また、第１方向特徴量と第２方向特徴量だけでもテクスチャの特徴量として意味があるが、その方向と周期より部分画像を切り出しテクスチャのパターン画像とする応用も考えられる。
【００４７】
＜画像特徴量抽出装置のテクスチャ比較検索処理＞
次にテクスチャ比較検索処理（Ｓ４）について説明する。ここでは、検索キーとなる画像（以下、キー画像と呼ぶ）をユーザーがキーボード３等を介して画面上より指定し、キー画像のテクスチャ特徴量とよく似たテクスチャ特徴量を持つ画像を類似性があると判断し画像データベース５０から検索し、ディスプレイ２等の表示部に結果を表示する処理を説明する。
【００４８】
図１２はテクスチャ比較検索処理のフローチャートである。最初に、ユーザーが指定したキー画像のテクスチャ情報（Ｔ0）を画像データベース５０からロードする（Ｓ４０１）。ここで、テクスチャ情報は、画像登録時に画像とともに画像データベース５０に登録されたものであり、ここでは、周期性フラグ、第１方向特徴量および第２方向特徴量等を含む。キー画像のテクスチャ情報が周期性を持つか否かを、テクスチャ特徴量抽出処理の過程で設定された周期性フラグにより判定する（Ｓ４０２）。周期性がなければ処理を終了し、周期性があればステップＳ４０３へ進む。
【００４９】
ステップＳ４０３では、検索対象の画像データベース５０においてレコードポインタｉをトップに移動する。レコードポインタiが示す画像のテクスチャ情報Ｔ[i]を画像データベース５０からロードする（Ｓ４０４）。
【００５０】
Ｓ４０４でロードしたテクスチャ情報Ｔ[i]の中の周期性フラグに基づいて周期性の有無を判断する（Ｓ４０５）。周期性がなければステップＳ４０９に進み、周期性があればステップＳ４０６に進む。
【００５１】
ステップＳ４０６では、キー画像のテクスチャ情報Ｔ0と検索対象のテクスチャ情報Ｔ[i]との類似度を比較演算し、数値化し、類似画像か否かを判定する（Ｓ４０６）。類似画像でなければ処理Ｓ４０９に進み、類似画像と判断すればステップＳ４０７に進む。ここで、類似度の計算は、第１周期特徴量における方向と周期および第２周期特徴量における方向と周期から計算してもよい。
【００５２】
例えば、
キー画像のテクスチャ情報Ｔ0の第一周期の周期をｄ₀（１）、
検索対象画像のテクスチャ情報Ｔ[ｉ]の第一周期の周期をｄ_i（１）、
キー画像のテクスチャ情報Ｔ0の第二周期の周期をｄ₀（２）、
検索対象画像のテクスチャ情報Ｔ[ｉ]の第二周期の周期をｄ_i（２）、
第一周期の方向の角度をθ₀、
第二周期の方向の角度をθ_i、
類似度判定のウェイトをＷ_A，Ｗθとすると、
周期が（ｄ₀（１），ｄ₀（２），θ₀）のテクスチャＴ0と、周期が（ｄ_i（１），ｄ_i（２），θ_i）のテクスチャＴiとの類似度Ｄは、次式で求められる。ここで、Ｄは大きい方が類似している。
【数２】

ここで、Ｒ0，Ｒiはパターン形状の縦横比であり、次式で表される。
Ｒ₀＝ｄ₀（２）／ｄ₀（１） …（８）
Ｒ_i＝ｄ_i（２）／ｄ_i（１） …（９）
【００５３】
ステップＳ４０７において、現在のレコードポインタｉが示す画像の情報をロードし（Ｓ４０７）し、このロードされた画像情報を検索結果とし、ディスプレイ２にサムネイル画像を表示する（Ｓ４０８）。検索結果表示の例としては、上記のようにサムネイル画像を表示する方法、画像の名前を表示する方法、サムネイルを生成した元画像のファイル名やパスを表示する方法、元画像の入手先を表示する方法などのバリエーションが考えられる。又は、類似性があると判定した結果の画像の数をカウントし、その数を表示する方法も考えられる。
【００５４】
その後、引き続き画像検索を行うために、レコードポインタｉを一つ進める（Ｓ４０９）。画像データベース５０中に検索対象の画像データがまだあるか否かを判定する（Ｓ４１０）。すなわち、レコードポインタｉを進めた結果、新たに読み出す画像データがまだあるか否かを判断する。検索対象の画像データが新たにないときは処理を終了し、検索対象の画像データがまだあるときはステップＳ４０４に戻り、上記処理（Ｓ４０４〜Ｓ４１０）を繰り返す。
【００５５】
以上のようにして、本実施形態の画像特徴量抽出装置は、画像の特徴を抽出する際に、画像データの一定幅の帯状部分の領域に対して濃度共起行列を求め、これに基づいて画像データの周期性（テクスチャ特徴量）を求めることにより、画像データのテクスチャ特徴量の抽出において処理速度を向上できる。
【００５６】
なお、本実施形態では、所定方向の帯上の領域の画像データに基づいてテクスチャ特徴量を検出するようにしたが、これに限らず、１本の線上の画像データでもよいし、所定間隔毎の数本の線上の画像データでもよい。但し、１本の線上の画像データで検出する場合、非常に高速であるが、扱える画像に制約がある。例えば、周期性を有することがある程度分かっている画像を扱うシステム（例えば、幾何学模様の布地データベース等）には適用可能である。また、所定間隔毎の数本の線上の画像データで検出する場合、少ないデータ量で検出ができ、かつ、比較的広範囲の領域に対して検出が可能である。
【００５７】
【発明の効果】
本発明の画像特徴量抽出装置は、画像データにおいて、一定幅の帯状部分の領域に対してテクスチャ特徴量を抽出するため、効率よく特徴量の抽出が行える。また、テクスチャ特徴量の抽出の際の前段処理であるデータの二値化処理において、その閾値を画素値分布における中間値より所定値だけシフトさせた値に設定することにより、画像中の大きな面積を占めかつ緩やかな諧調を持つ画像に対して元画像の特徴を損なうことなく画像の二値化が実現でき、画像の特徴量がより正確に抽出できる。
【図面の簡単な説明】
【図１】 本実施形態の画像特徴量抽出装置の構成の概略を示す図。
【図２】 画像特徴量抽出装置の制御装置を中心としたブロック図。
【図３】 画像データベースの構成図。
【図４】 濃度共起行列を求める際の走査方向θを示した図。
【図５】 各走査方向θにおいて、距離ｄに対して求められる慣性の値Ｉを示した図。
【図６】 画像特徴量抽出装置のメインのフローチャート。
【図７】 データ入力または設定等の画面の表示例を示した図。
【図８】 画像特徴量抽出装置におけるテクスチャ抽出処理のフローチャート。
【図９】 画像特徴量抽出装置における濃度共起行列計算処理のフローチャート。
【図１０】 画像特徴量抽出装置における特徴量抽出処理のフローチャート。
【図１１】 二値化処理の閾値を求める方法を説明した図。
【図１２】 画像特徴量抽出装置におけるテクスチャ比較検索処理のフローチャート。
【符号の説明】
１…制御装置、 ２…ディスプレイ、 ３…キーボード、 ４…マウス、 ５ａ…フロッピーディスク、 ５ｂ…フロッピーディスク装置、 ６…ハードディスク、 ７…プリンタ、 ８…スキャナ、 ９ａ…ＣＤ−ＲＯＭ、 ９ｂ…ＣＤ−ＲＯＭ装置、 １０…スピーカ、 １１…マイク、 ２１…初期設定画面、 ２３〜２５…選択用アイコン、 ２９，３０…設定用画面、 ５０…画像データベース、 ２０１…ＣＰＵ、 ２０２…クロック、 ２０３…ＲＯＭ、 ２０４…ＲＡＭ。[0001]
BACKGROUND OF THE INVENTION
The present invention extracts an image feature amount, and particularly relates to an image feature amount extraction apparatus that extracts a texture feature amount from an image.
[0002]
[Prior art]
In general, an image database stores and manages image data by adding attribute information such as creation date, modification date, file name, and file format, and search information such as keywords, shape characteristics, color, and sound. In such an image database, when data is registered, attribute information is generally automatically added. It is also generally known that feature quantities such as color, texture, and edge information are extracted from image data and added automatically. However, when extracting each feature amount from image data, the processing time for extraction becomes a problem. In particular, with regard to the texture feature amount, the amount of calculation of image processing for obtaining texture information from an image is large, and an efficient method has been demanded.
[0003]
[Problems to be solved by the invention]
For example, there is an invention disclosed in Japanese Patent Laid-Open No. 5-28266 as a method of extracting a texture feature amount from an image. In the present invention, the image is divided into square blocks, and the feature amount is calculated for each block. However, in this method, since the entire image is blocked and the feature amount is calculated, the processing target is the entire image, which is not efficient for an image having a uniform texture pattern.
[0004]
In general, image feature amount extraction is performed on image data after image normalization, graying, and binarization processing. Conventionally, in the binarization processing of an image, a binarization threshold value is set to an intermediate value or an average brightness value where the appearance probabilities of 0-value pixels and 1-value pixels are almost the same in the pixel value distribution after graying. Value was set. However, in this case, the pixel value of a portion that occupies a large area in the image and has a gentle gradation tends to be the threshold value, so the image including a background that looks uniform such as a blue sky or a plain background is binarized. There was a problem that the image at the time became unnatural. That is, in an image including a blue sky or a plain background, a portion of the background is converted to 0 and a portion of the background is converted to 1 even though it is the same background. There was a problem that it looked natural. Therefore, there is a problem that even if the feature amount of the image is extracted using such binarized data, the feature of the image cannot be appropriately extracted.
[0005]
The present invention has been made to solve the above problems, and an object of the present invention is to provide an image feature quantity extraction apparatus that efficiently extracts texture feature quantities from an image and more accurately extracts image features. There is to do.
[0006]
[Means for Solving the Problems]
  In order to achieve the above object, an image feature amount extraction apparatus according to the present invention extracts a texture feature amount from an image including a texture in which multi-valued pixel information is arranged in a two-dimensional grid. In the extraction device,Image input means for inputting an image, means for calculating a density co-occurrence matrix using pixels existing on a line in a predetermined direction or in an area of a certain width in the input image, and inertia based on the density co-occurrence matrix A means for calculating the presence or absence of periodicity of the image based on inertia, and a feature amount extraction means for extracting the texture feature amount of the image as the periodicity of the image when it is determined that there is periodicityHave
[0007]
  The determination means may determine that there is periodicity at a distance to which the inertia value is given when the minimum value of inertia is smaller than a predetermined value.
[0008]
In addition, the image feature amount extraction device includes an image input unit that inputs a grayed image, and an intermediate pixel value distribution in the grayed image based on the density of the pixels constituting the grayed image. Binarization conversion means for converting into a binarized image represented by 0 or 1 with a threshold set to a value shifted from the value by a predetermined value as a boundary, and binarized by the binarization conversion means Texture features are extracted from the captured images.
[0009]
  According to the present inventionimageThe feature quantity extraction method is an image feature quantity extraction method that extracts a texture feature quantity from an image including a texture in which multi-valued pixel information is arranged in a two-dimensional grid.An image is input, and in the input image, a density co-occurrence matrix is calculated using pixels existing on a line in a predetermined direction or in an area having a constant width, and inertia is calculated based on the density co-occurrence matrix. Based on whether the image has periodicity or not, the texture feature quantity of the image is extracted as the periodicity of the image when it is determined that the image has periodicity.
[0011]
  According to the present inventionRecordThe medium is a program for extracting a texture feature amount from an image including a texture in which multi-valued pixel information is arranged in a two-dimensional grid,A procedure for inputting an image, a procedure for calculating a density co-occurrence matrix using pixels existing on a line in a predetermined direction or in an area of a certain width in the input image, and calculating an inertia based on the density co-occurrence matrix And a computer for executing a procedure for determining the presence or absence of periodicity of the image based on inertia and a procedure for extracting the texture feature quantity of the image as the periodicity of the image when it is determined that there is periodicity.Record the program.
[0013]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of an image feature quantity extraction apparatus according to the present invention will be described below with reference to the accompanying drawings.
[0014]
The image feature amount extraction apparatus according to the present embodiment scans an image in a band-like region having a certain width, not the entire image, when extracting a feature amount of a texture from an image, and from an image in the region. Extract texture features. That is, the processing efficiency at the time of feature amount extraction is improved by extracting the image feature amount from a part of the image instead of the entire image.
[0015]
<Overall Configuration of Image Feature Extraction Device>
FIG. 1 shows a schematic configuration diagram of an image feature amount extraction apparatus (hereinafter referred to as “system”) of the present embodiment. As shown in FIG. 1, the system includes a central processing unit (hereinafter referred to as “CPU”), and is configured around a control device 1 that controls the entire system. As the CPU, for example, Pentium manufactured by Intel Corporation is used. The control device 1 includes a display 2 for displaying images or characters, a display for operation, a keyboard 3 and a mouse 4 for performing various inputs and instruction operations, and a floppy as a data storage medium. A disk device 5a and a hard disk device 6, a printer 7 for printing characters and image data, a scanner 8 for capturing image data, a CD-ROM device 9b for reading data stored in a CD-ROM 9a, A speaker 10 for outputting sound and a microphone 11 for inputting sound are connected.
[0016]
FIG. 2 shows a block diagram of this system. The CPU 201 is connected via a data bus 220 to a ROM 203 that stores a program for controlling the system and a RAM 204 that temporarily stores a program and data executed by the CPU 201 for control. Further, a circuit connected to the CPU 201 via the data bus 220 includes a display control circuit 205 that controls the display 2 for displaying images or characters, and a keyboard control circuit 206 that controls transfer of input from the keyboard 3. , A mouse control circuit 207 for transferring input from the mouse 4, a floppy disk device control circuit 208 for controlling the floppy disk device 5 b, a hard disk device control circuit 209 for controlling the hard disk device 6, and an output to the printer 7. A printer control circuit 210 for controlling, a scanner control circuit 211 for controlling the scanner 8, a CD-ROM device control circuit 212 for controlling the CD-ROM device 9 b, a speaker control circuit 213 for controlling the speaker 10, and the microphone 11. Microphone control circuit 214 to control There is. Further, a clock 202 for generating a reference clock necessary for operating the system is connected to the CPU 201, and an expansion slot 215 for connecting various expansion boards is connected via a data bus 220. . Note that an SCSII board may be connected to the expansion slot 215, and the floppy disk device 5b, the hard disk device 6, the scanner 8, the CD-ROM device 9b, or the like may be connected via the SCSII board.
[0017]
In the above system, the floppy disk 5a and the hard disk device 6 are used as the image data storage medium, but other information storage media such as a magneto-optical disk (MO) may be used. Further, although the scanner 8 is used as an image data input device, other data input devices such as a still video camera and a digital camera may be used. Further, although the printer 7 is used as an output device, other output devices such as a digital copying machine may be used. In this system, a program for realizing a data management system is stored in the ROM 203. However, a part or all of this program is stored in an information storage medium such as the floppy disk 5a, the hard disk device 6, or the CD-ROM 9b, and the program and data are read from the information storage medium to the RAM 204 as necessary. It may be executed.
[0018]
<Database and keyword dictionary used in image feature extraction device>
The system has an image database including data and additional information that serves as a search key for storing and managing image data. This image database is logically configured on an information storage medium such as the hard disk device 6. FIG. 3 shows an example of the configuration of the image database of this system. The database 50 shown in FIG. 3 includes “image data” that is image information to be managed, “keyword” that is one of search keys for the image data, and “feature amount 1” that is a feature amount that indicates the feature of the image data. ”,“ Feature amount 2 ”and the like.
[0019]
<About the concentration co-occurrence matrix>
Here, the density co-occurrence matrix and its inertia (inertia) used when extracting the image feature amount in this embodiment will be described. A method using a density co-occurrence matrix as a texture analysis method is a widely used method. This method is described in detail in, for example, “Fundamentals of Image Recognition [II] —Feature Extraction, Edge Detection, Texture Analysis—Shunji Mori et al.”.
[0020]
This method is basically based on the evaluation of the two-dimensional joint probability density function f (i, j | d, θ). f (i, j | d, θ) is a probability density function indicating that a pixel that is a distance d in the θ direction from a pixel having a density value i may have a density value j. A matrix of f (i, j | d, θ) for each (d, θ) is a density co-occurrence matrix, and i and j indicate the positions of rows and columns, respectively. If the texture is coarse and the distance d is small compared to the size of the texture components, a pair of pixels that are separated by (d, θ) will generally have similar density values, and thus the diagonal elements of the density co-occurrence matrix The nearby value becomes larger. On the other hand, if the distance d is as small as a texture component in a fine texture, a pair of pixels separated by (d, θ) may take various combinations of density values. It becomes relatively uniformly distributed over all elements.
[0021]
In the present embodiment, θ in the scanning direction when obtaining the density co-occurrence matrix is set to 0 °, 45 °, 90 °, and 135 ° as shown in FIGS. 4A to 4D (details will be described later). ). Here, the matrix S using the concentration co-occurrence matrix M (d, θ)_θ(D) is defined as follows.
S₀(D) = {M (d, 0 °) + M^t(D, 0 °)} / 2 (1)
S₄₅(D) = {M (d, 45 °) + M^t(D, 45 °)} / 2 (2)
S₉₀(D) = {M (d, 90 °) + M^t(D, 90 °)} / 2 (3)
S₁₃₅(D) = {M (d, 135 °) + M^t(D, 135 °)} / 2 (4)
Where M^t(D, θ) is a transposed matrix of M (d, θ).
By using these matrices, various feature quantities can be calculated, a feature space can be constructed, and textures can be identified.
[0022]
  Further, the inertia of the density co-occurrence matrix, which is one of the feature quantities, can be obtained by the following formula using the matrix.
[Expression 1]

here,S _θ ( i, j | D)Is a matrixS _θ (D)I row, j column element, N_GIs the number of density levels in the image.
[0023]
Since the periodicity of the texture is the periodicity of the density, the periodicity of the texture is S_θIt should be reflected in (i, j | d). Therefore, d is 3, 4,... D for each direction._maxThe presence / absence of periodicity is determined from the graph when changed (maximum value of distance d). This determination method will be described below. First, the inertia I {S for each distance d in each direction θ._θ(D)} for each inertia I {S_θ(D)} finds the minimum inertia value Iner (θ). Here, Iner (θ) is given by the following equation.

[0024]
Thereafter, the smallest one and the second smallest one are selected from the inner (0), inner (45), inner (90), and inner (135). If each selected minimum value Iner (θ) of inertia is smaller than a predetermined value, it is determined that the target texture has periodicity.
[0025]
For example, FIG. 5 shows the result of obtaining the inertia I when the distance d is changed with respect to a certain image in four directions θ. In this figure, the direction θ = 0 ° when the distance d = 7, the direction θ = 45 ° when the distance d = 7, and the direction θ = 90 ° when the distance d = 13. At 135 °, when the distance d = 6, the minimum value of inertia Iner (θ) in each direction is obtained. Of the minimum value Iner (θ) of the inertia I with respect to each direction θ, Inner (135) is the minimum, and Inner (45) is the second smallest value. Since both Inner (45) and Inner (135) are lower than the periodicity determination threshold, it is determined that this image is periodic in the directions of θ = 45 ° and 135 °.
[0026]
<Control Operation of Image Feature Extraction Device>
Below, the concrete control operation of this system is demonstrated using a flowchart.
[0027]
<Main flow>
FIG. 6 is a flowchart showing a main routine of a program executed in this system. When this program is started, first, initial setting processing for initializing flags and the like necessary for the subsequent processing and displaying an initial menu screen is performed (S1). FIG. 7 shows an example of the initial screen. On the initial menu screen 21, selection items 23 to 25 for selecting a predetermined process are displayed as icons, and when one of the selection items 23 to 25 is selected by the user, the predetermined process is executed. The In this system, on the setting screen such as the initial menu screen 21 displayed on the display 2 or the like, the user selects various processes and inputs setting values via the keyboard 3 or the mouse 4 or the like. Is called. After step S1, it is determined whether or not the user has made a menu selection on the initial menu screen 21 (S2). If “texture extraction” 23 is selected in step S 2, the process proceeds to a texture extraction process (S 3) in which a texture feature amount is extracted from the designated image data and registered in the image database 50. Proceed to step S6. In step S2, if “texture comparison search” 24 is selected, texture comparison is performed in which the designated image data is compared with the texture of the image data registered in the database 50, and similar image data is extracted. The process proceeds to the search process (S4), and then proceeds to step S6. If “other menu” 25 is selected in step S2, other menu processing is performed (S5), and then the process proceeds to step S6. If no menu is selected in step S2, the process proceeds to step S6 without doing anything. In step S6, other processes are executed. When all the processes are completed, the process returns to step S2, and thereafter the same process is repeated.
[0028]
Here, the other menu processing (S5) is basically the same as that of a general search system, and is not directly related to the present invention, so the description thereof will be omitted. In particular, texture extraction processing (step S3) The texture comparison search process (step S4) will be described in detail below.
[0029]
<Texture Extraction Processing in Image Feature Extraction Device>
First, the texture extraction process (step S3) will be described. FIG. 8 shows a flowchart of the texture extraction process. As shown in FIG. 8, in the texture extraction process, when registering image data in the image database 50, a density co-occurrence matrix calculation process (S31) for calculating a density co-occurrence matrix from the image data, and a density co-occurrence matrix calculation process Based on the result, a texture feature amount is extracted, and a feature amount extraction process (S32) for registering the feature amount together with the image data in the image database 50 is performed. Each step will be described below with reference to the flowcharts of FIGS.
[0030]
FIG. 9 is a flowchart of the density co-occurrence matrix calculation process (S31). In FIG. 9, first, an image is loaded from a storage medium such as a database or a file system (S301) and normalized (S302). The normalization here is to reduce or enlarge an acquired image to a predetermined size for the purpose of shortening the image processing time and reducing noise. In this embodiment, the image is reduced while maintaining the aspect ratio so that the image fits in the size of 120 × 120 pixels.
[0031]
Next, after the normalized data is converted into a gray image, binarization processing is performed (S303). Here, in this embodiment, in order to solve the problem of the binarization processing described in the related art, the threshold value is not set to a pixel value having a high appearance frequency, and values before and after the threshold value are set as the threshold value. That is, the threshold value is set to a value shifted forward and backward by a predetermined value from the intermediate value of the pixel value distribution after graying. This will be described below using a specific example.
[0032]
  FIG.Indicates the pixel value distribution with respect to the pixel value of the image after graying. In this figure, the pixel value has 256 gradations. In this embodiment,FIG.In the pixel value distribution, the threshold value is set to a value shifted from the intermediate value (M) of the pixel value to the higher pixel value by one third of the standard deviation (s). In this way, by setting the threshold value for binarization to a value shifted by a predetermined value from the intermediate value of the pixel value distribution, the threshold value is not set to the pixel value with high frequency, so that the characteristics of the original image are retained more. Binarization is possible. However, a threshold value different from the above is set in the following cases. That is, when the intermediate value M is 128 or more and the cumulative number of pixels at the intermediate value exceeds 90%, pixels whose pixel value is less than the intermediate value are set to “0”, and pixels whose intermediate value is equal to or higher than “ 1 ”. Further, when the intermediate value M is less than 128 and the cumulative number of pixels at the intermediate value exceeds 90%, the pixel having the pixel value equal to or lower than the intermediate value is set to “0”, and the pixel exceeding the intermediate value is set to “1”. " Further, when the cumulative number of pixels at a value of (intermediate value M + standard deviation s / 3) is 100%, the threshold value is set to the intermediate value M. By setting a threshold value and performing binarization processing in this way, a binarized image close to the original image is obtained as a result.
[0033]
When image normalization (S302) and binarization (S303) are completed, a density co-occurrence matrix is calculated from the image. First, the direction number n corresponding to the scanning direction θ for obtaining the density co-occurrence matrix is set to an initial value 0 (S304). The direction number n corresponds to a direction θ of 0 ° when 0, 45 ° when 1 and 90 ° when 2 and 90 ° when 3 and 135 ° when 3 respectively. Here, in order to obtain a more accurate directionality, for example, an angular interval for obtaining a concentration co-occurrence matrix is set every 10 °, and when the direction number is 0, 0 °, 1 is 10 °, and 2 is 20. The density co-occurrence matrix may be obtained by finely dividing the angle such as ° to correspond to the scanning direction θ. However, if the angle is divided finely, calculation processing increases and processing time is required. Therefore, here, the angle is divided by 45 °.
[0034]
Next, the pixel distance d when determining the density co-occurrence matrix is set to 1 (S305). A density co-occurrence matrix in the nth direction and distance d is obtained (S306). That is, in this case, the density co-occurrence matrix in the first direction and the distance 1 is obtained. In the present embodiment, when calculating the density co-occurrence matrix, it is not performed for the pixels of the entire image, but is performed for pixels on a band having a predetermined width in the same direction as the calculation target direction. The width of the band may be small enough to obtain the effect of speeding up the processing compared to the size of the image, and may be any size necessary for obtaining the density co-occurrence matrix. Naturally, if the width of the band is increased, the accuracy of the calculation result increases, but the number of pixels to be calculated increases, so the processing is slowed. Conversely, if the band is narrowed, the processing becomes faster but the accuracy of the calculation result decreases.
[0035]
In the present embodiment, considering that the normalized image size is 120 × 120, the band width will be described as 10 pixels. For example, in the case of the 0th direction, as shown in FIG. 4 (a), it passes through the center of the image and becomes a horizontal band having a width of 10 pixels. In the case of the first direction, as shown in FIG. 4B, the band passes through the center of the image and becomes a 10 ° wide 45 ° direction band. In the case of the second direction, as shown in FIG. 4C, the band passes through the center of the image and becomes a band of 90 ° direction with a width of 10 pixels. In the third direction, as shown in FIG. 4 (d), it passes through the center of the image and becomes a band of 135 ° direction with a width of 10 pixels. Note that the image area to be scanned does not have to be continuous as shown in FIG. 4 and may be an area composed of a plurality of discontinuous lines or areas as long as they have the same directionality. For example, you may scan the area | region which consists of 10 lines of 1 pixel located discretely. This is because if the entire image has periodicity, the periodicity can be obtained in this way as well.
[0036]
Using the equation (5), the inertia value I {S from the concentration co-occurrence matrix obtained in step S306._θ(D)} is obtained (S307). The inertia value obtained in step S307 is substituted into the array Val [d] (S308). Thereafter, the value of the distance d is incremented (S309).
[0037]
The set distance d is the upper limit d of the predetermined distance._maxIt is determined whether or not it exceeds (S310). Distance d is upper limit d_maxIf it exceeds, the process proceeds to S311. The distance d is the upper limit d_maxIf it is below, the process returns to step S306, and the distance d is the upper limit d._maxThe process (S306 to S310) is repeated until it exceeds. Where the upper limit of distance d_maxIs set to a value about half the size of the image.
[0038]
In step S310, from the value of inertia Val [d] for each distance d in the same direction, a minimum value of inertia Iner [n] in that direction is obtained (S311).
[0039]
Thereafter, in order to designate the next direction, the direction number n is incremented (S312). It is determined whether the incremented direction number n is 3 or less (S313). If the direction number n is 3 or less, the process returns to step S305, and the above process is repeated (S305 to S313). When the calculation of the inertia value is completed for all the direction numbers n, the process is terminated.
[0040]
As described above, when the inertia of the image with respect to a predetermined direction is obtained using the density co-occurrence matrix, the feature amount of the image is extracted based on the inertia value. Here, in this embodiment, the periodicity of the texture of the image is extracted as a texture feature amount. In general, it is well known that the inertia value obtained from the density co-occurrence matrix has a periodicity in the inertia value with respect to the distance d as the periodicity of the texture is noticeable when viewed in the same direction. . Then, the inertia value approaches 0 at a distance d that is a cycle. For this reason, in the present embodiment, a threshold for determining periodicity is determined, the threshold is compared with the minimum value of inertia, and the minimum value of inertia is given when the minimum value of inertia is smaller than the threshold. It is determined that there is periodicity at d.
[0041]
FIG. 10 is a flowchart of the feature amount extraction process (S32). In FIG. 10, the minimum value is selected from Inner [0] to Inner [3] storing the minimum value of inertia for each scanning direction (S314). Then, it is determined whether or not the minimum value is equal to or less than the threshold value for determining the periodicity (S315). If it exceeds the threshold value, the process proceeds to step S322, and if it is equal to or less than the threshold value, the process proceeds to step S316.
[0042]
In step S316, the direction number n and the period d for the minimum inertia determined to have periodicity are stored as texture feature quantities. Here, the periodicity that gives the minimum value of the minimum inertia is referred to as a “first period”. Further, the direction and period of the texture feature amount obtained here is referred to as a first periodic feature amount.
[0043]
Next, the second smallest inertia value is selected from the arrays Inner [0] to Inner [3] storing the minimum values of the respective inertias in the four scanning directions (S317). It is determined whether or not the value of inertia is equal to or less than a threshold value for determining that it has periodicity (S318), and if it exceeds the threshold value, the process proceeds to step S322. If it is equal to or less than the threshold value, the process proceeds to step S319.
[0044]
In step S319, the direction number n and the period d for the inertia having the second smallest value determined to have periodicity are stored as texture feature amounts. Here, the periodicity that gives the second smallest minimum inertia value is referred to as a “second period”. Further, the texture directionality and period obtained here are referred to as a second period feature amount.
[0045]
Thereafter, in step S320, the periodicity flag is set to “present”, and the process ends. Here, when the periodicity flag is “present”, it means that the target image has periodicity, and when it is “not present”, it means that the target image has no periodicity. In step S322, the periodicity flag is set to “none”. Finally, the image data is registered in the image database 50 together with the periodicity flag and the feature amount (when there is periodicity) (S321), and the process is terminated.
[0046]
In the above processing, information on whether or not the target image has periodicity is obtained by the periodicity flag. If there is further periodicity, the direction and period of the “first period” and the direction of “second period” And the period are obtained as texture features. Further, although only the first direction feature quantity and the second direction feature quantity are meaningful as texture feature quantities, an application in which a partial image is cut out based on the direction and period and used as a texture pattern image is also conceivable.
[0047]
<Texture comparison search processing of image feature extraction device>
Next, the texture comparison search process (S4) will be described. Here, the user designates an image as a search key (hereinafter referred to as a key image) on the screen via the keyboard 3 or the like, and an image having a texture feature amount similar to the texture feature amount of the key image is similar. A process of determining that there is a search, searching from the image database 50, and displaying the result on a display unit such as the display 2 will be described.
[0048]
  FIG.Is a flowchart of texture comparison search processing. First, the texture information (T0) of the key image designated by the user is loaded from the image database 50 (S401). Here, the texture information is registered in the image database 50 together with the image at the time of image registration, and here includes a periodicity flag, a first direction feature amount, a second direction feature amount, and the like. Whether the texture information of the key image has periodicity is determined by the periodicity flag set in the course of the texture feature amount extraction process (S402). If there is no periodicity, the process ends. If there is periodicity, the process proceeds to step S403.
[0049]
In step S403, the record pointer i is moved to the top in the image database 50 to be searched. The texture information T [i] of the image indicated by the record pointer i is loaded from the image database 50 (S404).
[0050]
The presence / absence of periodicity is determined based on the periodicity flag in the texture information T [i] loaded in S404 (S405). If there is no periodicity, the process proceeds to step S409, and if there is periodicity, the process proceeds to step S406.
[0051]
In step S406, the similarity between the texture information T0 of the key image and the texture information T [i] to be searched is compared and digitized to determine whether the image is a similar image (S406). If it is not a similar image, the process proceeds to step S409, and if it is determined that the image is similar, the process proceeds to step S407. Here, the similarity may be calculated from the direction and the period in the first periodic feature and the direction and the period in the second periodic feature.
[0052]
For example,
The period of the first period of the texture information T0 of the key image is d₀(1),
The period of the first period of the texture information T [i] of the search target image is d_i(1),
The period of the second period of the texture information T0 of the key image is d₀(2),
The period of the second period of the texture information T [i] of the search target image is d_i(2),
The angle in the direction of the first cycle is θ₀,
The angle in the direction of the second period is θ_i,
Weight of similarity determination is W_A, Wθ,
The period is (d₀(1), d₀(2), θ₀) Texture T0 and the period is (d_i(1), d_i(2), θ_iThe degree of similarity D with the texture Ti is obtained by the following equation. Here, the larger D is similar.
[Expression 2]

Here, R0 and Ri are aspect ratios of the pattern shape and are expressed by the following equations.
R₀= D₀(2) / d₀(1) ... (8)
R_i= D_i(2) / d_i(1) ... (9)
[0053]
In step S407, information on the image indicated by the current record pointer i is loaded (S407), the loaded image information is used as a search result, and a thumbnail image is displayed on the display 2 (S408). Examples of search result display include displaying the thumbnail image as described above, displaying the name of the image, displaying the file name and path of the original image that generated the thumbnail, and displaying the source of the original image. Variations such as how to do is conceivable. Alternatively, a method of counting the number of images determined as having similarity and displaying the number is also conceivable.
[0054]
Thereafter, the record pointer i is advanced by one to continue the image search (S409). It is determined whether there is still image data to be searched in the image database 50 (S410). That is, as a result of the advancement of the record pointer i, it is determined whether there is still image data to be newly read. If there is no new search target image data, the process ends. If there is still search target image data, the process returns to step S404, and the above processes (S404 to S410) are repeated.
[0055]
As described above, the image feature amount extraction apparatus according to the present embodiment obtains a density co-occurrence matrix for a region of a band-like portion having a certain width of image data when extracting image features, and based on this By obtaining the periodicity (texture feature amount) of the image data, the processing speed can be improved in extracting the texture feature amount of the image data.
[0056]
In the present embodiment, the texture feature amount is detected based on the image data of the area on the band in the predetermined direction. However, the present invention is not limited to this, and the image data on one line may be detected or at predetermined intervals. The image data on several lines may be used. However, when detection is performed using image data on one line, it is very fast, but there are restrictions on the images that can be handled. For example, the present invention can be applied to a system that handles an image that is known to have a certain degree of periodicity (for example, a textile database of geometric patterns). Further, when detection is performed using image data on several lines at predetermined intervals, detection can be performed with a small amount of data, and detection can be performed over a relatively wide area.
[0057]
【The invention's effect】
The image feature amount extraction apparatus according to the present invention extracts a texture feature amount from an area of a band-shaped portion having a certain width in image data, so that the feature amount can be extracted efficiently. In addition, in the binarization processing of data, which is the pre-stage processing at the time of texture feature extraction, by setting the threshold value to a value shifted by a predetermined value from the intermediate value in the pixel value distribution, a large area in the image The image can be binarized without losing the characteristics of the original image with respect to the image that occupies the image and has a gentle gradation, and the feature amount of the image can be extracted more accurately.
[Brief description of the drawings]
FIG. 1 is a diagram showing an outline of a configuration of an image feature amount extraction apparatus according to an embodiment.
FIG. 2 is a block diagram centering on a control device of the image feature quantity extraction device;
FIG. 3 is a configuration diagram of an image database.
FIG. 4 is a diagram showing a scanning direction θ when obtaining a density co-occurrence matrix.
FIG. 5 is a diagram illustrating an inertia value I obtained with respect to a distance d in each scanning direction θ.
FIG. 6 is a main flowchart of the image feature quantity extraction device.
FIG. 7 is a diagram showing a display example of a screen for data input or setting.
FIG. 8 is a flowchart of texture extraction processing in the image feature quantity extraction device;
FIG. 9 is a flowchart of density co-occurrence matrix calculation processing in the image feature quantity extraction device.
FIG. 10 is a flowchart of feature amount extraction processing in the image feature amount extraction apparatus;
FIG. 11 is a diagram illustrating a method for obtaining a threshold value for binarization processing;
FIG. 12 is a flowchart of texture comparison search processing in the image feature quantity extraction device.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... Control apparatus, 2 ... Display, 3 ... Keyboard, 4 ... Mouse, 5a ... Floppy disk, 5b ... Floppy disk device, 6 ... Hard disk, 7 ... Printer, 8 ... Scanner, 9a ... CD-ROM, 9b ... CD- ROM device, 10 ... speaker, 11 ... microphone, 21 ... initial setting screen, 23-25 ... selection icon, 29, 30 ... setting screen, 50 ... image database, 201 ... CPU, 202 ... clock, 203 ... ROM, 204: RAM.

Claims (6)

In an image feature amount extraction device that extracts a texture feature amount from an image including a texture in which multi-valued pixel information is arranged in a two-dimensional grid,
An image input means for inputting an image;
Means for calculating a density co-occurrence matrix using pixels existing in a line in a predetermined direction or in an area having a constant width in the input image;
Means for calculating inertia based on the concentration co-occurrence matrix;
Determining means for determining the presence or absence of periodicity of the image based on the inertia;
An image feature quantity extraction device comprising: feature quantity extraction means for extracting a texture feature quantity of an image as the periodicity of an image when it is determined that the image has periodicity . The image feature amount extraction device according to claim 1,
  An image feature amount extraction apparatus, wherein the determination unit determines that there is periodicity at a distance to which the inertia value is given when the minimum value of inertia is smaller than a predetermined value.   2. The image feature amount extraction apparatus according to claim 1, wherein the pixel value in the image after graying is based on the image input means for inputting the grayed image and the density of the pixels constituting the grayed image. Binarization conversion means for converting into a binarized image represented by 0 or 1 with a threshold set to a value shifted by a predetermined value from an intermediate value of the distribution as a boundary, and binarization conversion means by the binarization conversion means An image feature quantity extraction device that extracts a texture feature quantity from a digitized image. In an image feature amount extraction method for extracting a texture feature amount from an image including a texture in which multi-valued pixel information is arranged in a two-dimensional grid,
Enter an image,
In the input image, calculate the density co-occurrence matrix using pixels that exist on a line in a predetermined direction or in an area of a certain width,
Calculating inertia based on the concentration co-occurrence matrix;
Determine the presence or absence of periodicity of the image based on the inertia,
An image feature amount extraction method, wherein a texture feature amount of an image is extracted as the periodicity of the image when it is determined that the periodicity is present . A program for extracting a texture feature amount from an image including a texture in which multi-valued pixel information is arranged in a two-dimensional grid,
The procedure for entering images,
In the input image, a procedure for calculating a density co-occurrence matrix using pixels existing on a line in a predetermined direction or in an area having a constant width;
Calculating inertia based on the concentration co-occurrence matrix;
A procedure for determining the presence or absence of periodicity of the image based on the inertia;
A computer-readable recording medium having recorded thereon a program for causing a computer to execute a procedure of extracting a texture feature amount of an image as a periodicity of an image when it is determined that the image has periodicity . In the image feature amount extraction device according to any one of claims 1 to 3,
An image specifying means for specifying an image as a search key;
Similarity calculation means for calculating the similarity between both images based on the periodicity as the texture feature amount for the search key image and the periodicity as the texture feature amount for the search target image ;
Determination means for determining whether or not both images are similar based on the similarity by the similarity calculation means;
An image feature quantity extraction device further comprising: an image output means for outputting the search target image as a search result when both images are similar based on a determination result by the similarity calculation means. .

Images