JP3826412B2 - Edge detection method and edge detection apparatus - Google Patents

Description

【００５８】
Ｎ ＝ ＲＴ
となる。このようにして求められた法線ベクトルＮは、スプライン曲線Ｐ（ｔ）の法線ベクトルであるため、エッジの向きと推定した輪郭の方向が平行であるならば、各々に垂直な後述するグラディエントとスプライン曲線Ｐ（ｔ）の法線ベクトルＮも平行となる。
【００５９】
次に、上述したグラディエント計算部１１は、スプライン曲線発生部１２からの対象画素（ｘ，ｙ）の近傍３×３画素の領域について、ステップＳ１の処理において取り込んだ入力画像Ｆ_inの画素値Ｉ（ｘ，ｙ）を読み込み、対象画素（ｘ，ｙ）における画像の勾配（以下、グラディエントと言う。）Ｇを求める。
【００６０】
すなわち、グラディエント計算部１１において、上記図２に示すように、フィルタ演算器１１１は、読み込んだ入力画像Ｆ_inの画素値Ｉ（ｘ，ｙ）に対してＸ方向のソーベルフィルタＳ_Xを施す。また、フィルタ演算器１１２は、読み込んだ入力画像Ｆ_inの画素値Ｉ（ｘ，ｙ）に対してＹ方向のソーベルフィルタＳ_Yを施す。そして、グラディエント計算部１１は、フィルタ演算器１１１及びフィルタ演算器１１２の出力Ｄｘ，Ｄｙを組み合わせて、対象画素（ｘ，ｙ）における画像のグラディエントＧ（＝（Ｄｘ，Ｄｙ））としてエッジ強度計算部１４に供給する（ステップＳ４．２．２）。
【００６１】
上述のようにして得られたグラディエントＧは、画像の濃淡を標高としてとらえた場合に、傾斜が最大の方向を示すベクトルであり、エッジの方向に垂直なものである。すなわち、グラディエントＧは、入力画像Ｆinの画素値Ｉ（ｘ，ｙ）に対して、Ｇ ＝ ｇｒａｄ Ｉ（ｘ，ｙ） ＝ （Ｄｘ，Ｄｙ）
なる式で表すことができる。
【００６２】
次に、エッジ強度計算部１４は、グラディエント計算部１１で得られたグラディエントＧの長さｌｅｎＧと、グラディエントＧの長さｌｅｎＧが「１」となるように正規化したｎｏｒｍＧとを求める（ステップＳ４．２．３）。
ここで、エッジ強度計算部１４は、上記図３に示すように、グラディエントテーブルＴ_Gを備えている。このグラディエントテーブルＴ_Gには、グラディエントＧの長さｌｅｎＧと、それを正規化したｎｏｒｍＧとが［数２］に示す演算式を用いて予め計算された結果が格納されている。
【００６３】
【数２】

【００６４】
具体的に説明すると、グラディエントテーブルＴ_Gは、グラディエント計算部１１で得られたグラディエントＧ（＝（Ｄｘ，Ｄｙ））を入力とし、グラディエントＧの長さｌｅｎＧと、それを正規化したｎｏｒｍＧとを出力とするテーブルである。また、画素値の精度は、通常８ビットであり、Ｘ方向の出力Ｄｘ及びＹ方向の出力Ｄｙも同じ精度が使用されるため、グラディエントテーブルＴ_Gには、Ｘ方向の出力ＤｘとＹ方向の出力Ｄｙの合わせて１６ビットが入力されるようになされている。このため、任意のＸ方向の出力ＤｘとＹ方向の出力Ｄｙの組み合わせに対して、グラディエントテーブルＴ_Gの１要素が割り当てられる。したがって、グラディエントテーブルＴ_Gの各要素に各々、グラディエントＧの長さｌｅｎＧの値と、それを正規化したｎｏｒｍＧの値とが格納されている。このようなグラディエントテーブルＴ_Gを用いることにより、平方根等の複雑な計算を行うこと無しに、容易に結果を得ることができる。また、グラディエントテーブルＴ_Gへの入力は、１６ビットとしているため、テーブルサイズ、すなわち要素の数は、「６５５３６」であり、現在の計算機に十分適用することができる。
【００６５】
上述のようなグラディエントテーブルＴ_Gにより、グラディエント計算部１１で得られたグラディエントＧ（＝（Ｄｘ，Ｄｙ））のグラディエントＧの長さｌｅｎＧと、それを正規化したｎｏｒｍＧが求められる。そして、求められたグラディエントＧの長さｌｅｎＧを正規化したｎｏｒｍＧは、内積計算部１４１に供給されると共に、グラディエントＧの長さｌｅｎＧは、乗算器１４３に供給される。
【００６６】
最後に、エッジ強度計算部１４は、対象画素（ｘ，ｙ）における輪郭のエッジ強度Ｅ（ｘ，ｙ）を上述のようにして求めた法線ベクトルＮ、グラディエントテーブルＴ_GからのグラディエントＧの長さｌｅｎＧ、それを正規化したｎｏｒｍＧ、及び方向選択指数ｓを持って、
Ｅ（ｘ，ｙ） ＝ ｌｅｎＧ（ｎｏｒｍＧ・Ｎ）^S
なる式により求める。
【００６７】
すなわち、エッジ強度計算部１４において、上記図３に示すように、先ず、内積計算部１４１は、法線ベクトルＮと、グラディエントＧの長さｌｅｎＧを正規化したｎｏｒｍＧとの内積「ｎｏｒｍＧ・Ｎ」を、乗算により求める。この結果がグラディエントＧと法線ベクトルＮのなす角度θのコサイン「ｃｏｓθ」となる。
【００６８】
次に、コサインテーブルＴ_Cを用いて、供給される方向選択指数ｓにより、
ｃｏｓ^Sθ
を求める。このコサインテーブルＴ_Cは、内積計算部１４１で得られたコサイン「ｃｏｓθ」を入力とし、コサイン「ｃｏｓθ」のＳ乗を出力とするテーブルであり、コサインテーブルＴ_Cから出力されるコサイン「ｃｏｓθ」のＳ乗（＝ｃｏｓ^Sθ）は、乗算器１４３に供給される。ここで、本実施例では、方向選択指数ｓは、予め設定されている固定値としているため、コサインテーブルＴ_Cへの入力は、内積計算部１４１で得られた内積の結果であり、コサインテーブルＴ_Cは、入力が８ビット程度の小さなテーブルである。このようなコサインテーブルＴ_Cを用いることにより、複雑な計算を行うこと無しに、コサインのＳ乗「ｃｏｓ^Sθ」を容易に求めることができる。
【００６９】
そして、最後に、乗算器１４３は、グラディエントテーブルＴ_GからのグラディエントＧの長さｌｅｎＧと、コサインテーブルＴ_CからのコサインのＳ乗「ｃｏｓ^Sθ」とを掛け合わせ、掛け合わせた結果をエッジ強度Ｅとして出力する。ここで、エッジ強度計算部１４に供給される方向選択指数ｓは、どの程度選択的に特定方向のエッジにだけ反応するかを決定する指数である。この方向選択指数ｓは、値が大きければその分指向性が強くなるため、不要なエッジの混入をより強く阻止することができるが、所望の方向、すなわち推定したエッジ方向と実際のエッジ方向が少しでもずれていた場合には、エッジ強度の減衰を引き起こす。そこで、方向選択指数ｓは、通常２〜８の間の値が設定されている。
【００７０】
例えば、図６に示すように、方向選択指数ｓが大きくなると、グラディエントＧと法線ベクトルＮが同じ方向、すなわち内積「ｎｏｒｍＧ・Ｎ」の値が「１」に近い場合のみ、エッジ強度Ｅ（ｘ，ｙ）が大きい値となる。すなわち、高い指向性が得られる。また、方向選択指数ｓを「１」とした場合には、上述したロビンソンの３レベルのフィルタ、プレウイットのコンパスフィルタ、及びキルシュのフィルタにおける指向性と同じ指向性が得られる。
【００７１】
上述のようなステップＳ４の処理を概略の輪郭を与えるスプライン曲線Ｐ（ｔ）上の各画素（ｘ，ｙ）に対して行う。したがって、エッジ強度計算部１４からは、図７に示すように、入力画像Ｆ_inから得られたエッジ強度画像Ｆ_outが出力される（ステップＳ５）。
上述のように、本実施例では、エッジ強度Ｅを求める処理において、グラディエントＧと法線ベクトルＮのなす角度を直接求めずに、内積の計算により直接コサイン「ｃｏｓθ」を求めているため、３角関数の計算等の複雑な処理を行う必要が無い。すなわち、グラディエントＧと法線ベクトルＮのなす角度の大きさと対応する量が、乗算２回と加算１回のみのはるかに少ない計算量で算出することができる。
【００７２】
また、従来のエッジ検出方法及びエッジ検出装置では、上記図１２の（ｂ）に示した前景物体内の色の差や背景物体の交差による不要なエッジ２２１ａ〜２２１ｃの混入が不可避であり、例えば、推定したエッジ方向から４５°ずれた方向のエッジの強度は２９％の減衰、６０°ずれた方向のエッジの強度は５０％の減衰しか得られないのに対して、本実施例によると、上記図６に示すように、方向選択指数ｓを「４」とした場合、推定したエッジ方向から４５°ずれた方向のエッジの強度は７５％の減衰、６０°ずれた方向のエッジの強度は９４％の減衰を得ることができる。また、上述のような前景内の色変化や背景によって生じるうるエッジ２２１ａ〜２２１ｃの方向は、対象物体の輪郭の方向とは大きく異なる角度となるため、上述したような強い指向性を用いることにより、不要なエッジの混入をほぼ完全に除去することができる。
【００７３】
また、概略の輪郭の方向の推定値である接線ベクトルＴ、及び入力画像Ｆ_inのグラディエントＧを輪郭上の各画素毎に求めるため、輪郭上の各点において輪郭に沿ったエッジを選択的に検出することができる。
例えば、上記図５（ａ）に示した画像２５０において、円形の物体２６０内の色の違いがある箇所Ａ，Ｂでは、法線ベクトルＮは、曲線２７０に垂直な方向となり、円形の物体２６０内の色の違いにより生じる水平な線上では、グラディエントＧは、上記水平な線に対して垂直方向となる。このため、上記法線ベクトルＮと上記グラディエントＧの角度差は、９０゜となり、上記図６に示すように、エッジ強度Ｅは、「０」となる。すなわち、円形の物体２６０内のエッジは、検出されない。一方、輪郭を表すエッジは、曲線２７０に沿うため、箇所Ａ，ＢにおけるグラディエントＧは、水平方向となる。これは法線ベクトルＮと同じ方向であり、このため、輪郭を表すエッジは、強く検出される。したがって、上記図１２（ｂ）に示したような不要なエッジ２２１ａ，２２１ｂ，２２１ｃを混入することなく、本来求めたいエッジのみを検出することができる。
【００７４】
また、エッジ強度計算部１４に対してエッジの方向を推定した情報である接線ベクトルＴを補助情報として与え、エッジ強度計算部１４を特定の方向に強く反応するものとすることにより、不要なエッジが混入することなく必要なエッジのみを高いＳ／Ｎ比で検出することができる。
【００７５】
また、オペレータにより入力された輪郭の概略から各位置での対象物体のエッジの方向を算出できるため、オペレータの作業量を増やすこと無く、手間をかけずに必要なエッジのみを検出することができる。
尚、上述した実施例では、方向選択指数ｓの設定値を２〜８の間の値としたが、エッジの推定方向の精度をさらに高くし、方向選択指数ｓにさらに大きな値を設定してもよい。これにより、指向性をさらに強めることができる。
【００７６】
また、方向選択指数ｓの値を外部から入力することにより、状況に応じて方向選択指数ｓの値を変化させてもよい。通常、方向選択指数ｓの値は、数通り選択できれば十分であるため、方向選択指数ｓを入力とするコサインテーブルＴ_C追加ビット数は、４ビット程度でよい。また、方向選択指数ｓの値を整数に限定して、計算機が乗算の繰り返しとして冪乗を計算してもよい。
【００７７】
また、上述した実施例では、コサイン「ｃｏｓθ」の冪乗を使用してエッジ強度を求めることとしたが、上記図３に示した内積演算部１４１の出力をＩＰ、指向性の分布関数をｆ（θ）として、
ｆ（ｃｏｓ^-1ＩＰ）
を計算し、その結果を使用してエッジ強度を求めてもよい。これにより、例えば、図８に示すように、推定したエッジ方向と実際のエッジ方向のなす角度θがしきい値Ｋ以下の場合には、エッジ強度を減衰させず、しきい値Ｋを超えた場合には、エッジ強度を急峻に減衰させる指向性を得ることができる。
【００７８】
具体的に説明すると、先ず、所望の分布形状を定義し、上記図８に示したグラフの横軸の値をコサインに変換する。次に、コサインの値各々に対する減衰率、すなわち「ｆ（ｃｏｓ^-1ＩＰ）」の計算結果をテーブルに格納する。このテーブルを上述した実施例で用いたコサインテーブルＴ_Cの代わりに使用する。したがって、分布関数ｆ（θ）を使用してエッジ強度を求めることができる。
【００７９】
また、分布関数ｆ（θ）が方向選択指数ｓをも変数として持つこととしてもよい。すなわち、分布関数ｆ（θｓ）を使用してエッジ強度を求めてもよい。この場合には、方向選択指数ｓをも上記テーブルの入力とし、方向選択指数ｓを変化させながらエッジ強度を求める。
【００８０】
また、上述した実施例では、輪郭の概略を３次スプライン形式で表現されるスプライン曲線Ｐ（ｔ）としたが、異なる次数のスプライン形式で表現されるものとしてもよい。また、助変数ｔを用いずに、インプリサイトフォーム（ｉｍｐｌｉｃｉｔ ｆｏｒｍ）形式で表現される曲線
Ｑ（ｘ，ｙ） ＝ ０
としてもよい。
【００８１】
また、オペレータが描画した概略の輪郭線の画像を細線化し、隣接する画素の位置関係からエッジの方向を推定してもよい。この場合には、上記図４に示したステップＳ３において、概略の輪郭線の画像を細線化し、上記図４に示したステップＳ４．１において、細線化した画像中の対象画素に隣接する複数の画素の並びからその位置における輪郭の方向を求めるようにする。
【００８２】
また、上記図４に示したステップＳ４において、特定の方向のエッジに選択的に反応するような複数のフィルタ、例えば、コンパスオペレータ（ｃｏｍｐａｓｓ ｏｐｅｒａｔｏｒ）を用いて、エッジの推定した方向を基にしてその方向に対する検出能力の最も高いフィルタを上記複数のフィルタから選択してエッジ強度を求めるようにしてもよい。
【００８３】
また、上述した実施例では、輪郭の概略を表す情報がオペレータにより入力されるものとしたが、複数枚の時間的に連続する画像、すなわち動画像から連続的にエッジを検出することとしてもよい。
この場合には、上記図４に示したフローチャートのステップＳ２の処理において、概略の輪郭を座標列（ｘｎ，ｙｎ）として取り込むのではなく、現在の処理対象の画像に対して直前の画像、直前の画像から得られたエッジ検出の結果、及び現在の処理対象の画像の３つの情報から現在の処理対象の画像における輪郭を推定する。
【００８４】
すなわち、図９に示すように、先ず、現在の処理対象の画像Ｆ_Cの直前の画像Ｆ_Pに対するエッジ検出処理の結果得られたエッジ３０上に一定間隔で画素（以下、マークと言う）３１₁，３１₂，３１₃，・・・を結ぶ。次に、各マーク３１₁，３１₂，３１₃，・・・に対して、各々マークを含む小ブロックを定める。例えば、マーク３１₁に対しては、マーク３１₁を含む小ブロック３２を定める。このようにして定めた小ブロック内のデータと最も相関が高い部分を現在の処理対象の画像Ｆ_C中からブロックマッチング処理により検出する。
【００８５】
このブロックマッチング処理においては、先ず、例えば、小ブロック３２内の画素値と、現在の処理対象の画像Ｆ_C中の同じ大きさのブロック内の画素値とを比較し、その差の総和の大小により相関の高さを決定することにより、最も相関が高いブロック３３を検出する。この時、ブロック３３における画素３４が小ブロック３２のマーク３１₁に対応する画素となる。マーク３１₁以外のマーク３１₂，３１₃，・・・に対しても同様にして、対応する画素を現在の処理対象の画像Ｆ_C中から検出する。
【００８６】
したがって、上記図４に示したフローチャートのステップＳ３の処理において、上述のようにして得られた一連の対応する画素の座標列を用いて、スプライン曲線Ｐ（ｔ）を生成する。
上述のように、現在の処理対象の画像と直前の画像からブロックマッチング処理により連続的に輪郭を抽出し、その輪郭を用いてエッジ検出を行うことにより、動画像に対してオペレータの介在なく自動的にエッジを検出することができる。また、この場合、ブロックマッチング処理に用いる画像を現在の処理対象の画像と直前の画像のみではなく、画像の枚数を増やしてブロックマッチング処理を行ってもよい。これにより、ブロックの移動量の算出精度を高めることができるため、より正確にエッジを検出することができる。
【００８７】
また、時間的に離れた複数の画像において、上記図４に示したフローチャートのステップＳ２の処理にキーフレーム法を適用してもよい。すなわち、複数の画像についてオペレータが形状の指示を与え、中間の画像については計算機が補間によりオペレータの代替を行う手法を適用してもよい。この場合には、上記ステップＳ２の処理において、概略の輪郭をオペレータが設定し、中間の画像群については、概略の輪郭上の点群を補間し、自動的にその画像におけるスプライン曲線を生成する。
【００８８】
上述のように複数枚の画像と過去のエッジ検出の結果を基にして現在の処理対象の画像におけるエッジを検出することにより、フレーム間の相関を用いてさらに精度良くエッジ方向又は色の変化の情報である補助情報を求めることができる。したがって、エッジ検出精度をさらに高めることができ、動画像におけるエッジ検出を手間をかけずに行うことができる。
【００８９】
【発明の効果】
本発明に係るエッジ検出方法では、推定手段によりあらかじめ示されているエッジ抽出処理対象の点群に沿うように曲線を発生させ、発生させた曲線上に存在する点に対して、入力画像データにおける各位置のエッジ方向を推定し、グラディエント算出手段により入力画像データにおける各位置のグラディエントを求め、グラディエント強度算出手段により、上記推定したエッジ方向と、上記グラディエントのなす角度と、上記グラディエントの大きさと、上記推定したエッジ方向と上記グラディエントとのなす角度の絶対値が小さいほど大きい値をとる指向性を決定する指数と、からエッジ強度を算出してエッジを検出する。
これにより、エッジ検出において強い指向性を得ることができ、不要なエッジが混入すること無く、対象物体の輪郭を構成するエッジのみを検出することができる。したがって、検出精度を高めることができる。
【００９０】
また、本発明に係るエッジ検出方法では、上記推定したエッジ方向と上記グラディエントのなす角度のコサインを求めてエッジ強度を算出する。これにより、少ない計算量でエッジ強度を算出することができる。したがって、良好なエッジ画像を容易に得ることができる。
【００９１】
また、本発明に係るエッジ検出方法では、上記推定したエッジ方向と上記グラディエントのなす角度と、上記グラディエントの大きさと、指向性を決定する指数から、エッジ強度を算出するので、少ない計算量でエッジ強度を算出することができる。したがって、良好なエッジ画像を容易に得ることができる。
【００９２】
また、本発明に係るエッジ検出方法では、上記推定したエッジ方向と上記グラディエントのなす角度のコサインを、予め正規化したベクトルである上記推定したエッジ方向と、長さを「１」に正規化した上記グラディエントとの内積として求める。これにより、さらに強い指向性を得ることができ、さらに少ない計算量でエッジ強度を算出することができる。したがって、良好なエッジ画像をさらに容易に得ることができる。
【００９４】
本発明に係るエッジ検出装置では、あらかじめ示されているエッジ抽出処理対象の点群に沿うように曲線を発生させ、発生させた曲線上に存在する点に対して、入力画像データにおける各位置のエッジ方向を推定手段により推定するとともに、入力画像データにおける各位置のグラディエントをグラディエント算出手段により求め、上記推定手段で推定したエッジ方向と、上記グラディエント算出手段で得られたグラディエントのなす角度と、上記グラディエントの大きさと、指向性を決定する指数から、エッジ強度算出手段によりエッジ強度を算出し、上記エッジ強度算出手段で得られたエッジ強度から検出手段によりエッジを検出する。上記指数は、上記推定したエッジ方向と上記グラディエントのなす角度の絶対値が小さいほど、大きい値をとる。
これにより、エッジ検出において強い指向性を得ることができ、不要なエッジが混入すること無く、対象物体の輪郭を構成するエッジのみを検出することができる。したがって、エッジ検出精度を高めることができる。
【００９５】
また、本発明に係るエッジ検出装置では、上記エッジ強度算出手段は、上記推定したエッジ方向と上記グラディエントのなす角度のコサインを求めてエッジ強度を算出する。これにより、少ない計算量でエッジ強度を算出することができる。したがって、良好なエッジ画像を容易に得ることができる。
【００９６】
また、本発明に係るエッジ検出装置では、上記エッジ強度算出手段により、上記推定したエッジ方向と上記グラディエントのなす角度と、上記グラディエントの大きさと、指向性を決定する指数から、エッジ強度を算出するので、少ない計算量でエッジ強度を算出することができる。したがって、良好なエッジ画像を容易に得ることができる。
【００９７】
また、本発明に係るエッジ検出装置では、上記エッジ強度算出手段は、上記推定したエッジ方向と上記グラディエントのなす角度のコサインを、予め正規化したベクトルである上記推定したエッジ方向と、長さを「１」に正規化した上記グラディエントとの内積として求める。これにより、さらに強い指向性を得ることができ、さらに少ない計算量でエッジ強度を算出することができる。したがって、良好なエッジ画像をさらに容易に得ることができる。
【図面の簡単な説明】
【図１】本発明に係るエッジ検出装置の構成を示すブロック図である。
【図２】上記エッジ検出装置のグラディエント計算部の構成を示すブロック図である。
【図３】上記エッジ検出装置のエッジ強度計算部の構成を示すブロック図である。
【図４】上記エッジ検出装置におけるエッジ検出処理を示すフローチャートである。
【図５】概略の輪郭を与える曲線を説明するための図である。
【図６】方向選択指数とエッジ強度の関係を示すグラフである。
【図７】入力画像とエッジ強度画像の関係を説明するための図である。
【図８】指向性分布関数を用いてエッジ強度を求める場合を説明するための図である。
【図９】ブロックマッチング処理を説明するための図である。
【図１０】特定方向のエッジを検出するソーベルフィルタを示す略線図である。
【図１１】従来のエッジ検出方法を用いて物体内部に色の違いがある画像のエッジ検出を行った場合の結果を説明するための図である。
【図１２】オペレータにより入力される輪郭の概略を使用する従来のエッジ検出方法を用いて物体内部に色の違いがある画像のエッジ検出を行った場合の結果を説明するための図である。
【図１３】特定方向のエッジを検出するロビンソン３レベルフィルタ群を示す略線図である。
【図１４】従来のエッジ検出方法における指向性と実用上必要となる指向性の比較を説明するための図である。
【符号の説明】
１０ エッジ検出装置
１１ グラディエント計算部
１２ スプライン曲線発生部
１３ パラメータ座標生成部
１４ エッジ強度計算部[0001]
[Industrial application fields]
The present invention relates to an edge detection method and an edge detection apparatus that play a fundamental role in image processing, and special effect processing in video production such as television and movies, and component recognition from camera images in FA (Factory Automation). The present invention relates to an edge detection method and an edge detection apparatus that are suitable for application to the above.
[0002]
[Prior art]
The operation of cutting out an arbitrary partial image from an image is a basic operation for generating texture and structure data in video editing and synthesis, computer graphics, and the like. This operation of cutting out requires processing such as edge detection, region extraction, and association with a known object. In particular, when the background or the target object in the image becomes complicated, it is necessary to accurately obtain and track the edges constituting the contour of the target object.
[0003]
Here, the edge detection is a process of finding a portion where the pixel value changes sharply in the grayscale image. Usually, since the steep change occurs in the contour of the target object, the contour of the target object can be extracted from the image based on the result of edge detection. Therefore, edge detection is used in various fields as the most basic processing for obtaining information about a target object existing therein from an image. In particular, the extraction of the contour of the target object is the most basic process for determining the configuration, position, and the like of the target object existing in the image by a computer or the like, and is the most important application for edge detection.
[0004]
There are many known methods for detecting an edge as described above. However, any of these methods examines a local change in the value of a pixel and detects a portion where the change is large as an edge. As a typical edge detection method, there is a method using a differential filter as a spatial filter. This edge detection method using a differential filter is a method for detecting an edge based on the principle that the differential coefficient also has a large value in a portion where the change in the pixel value is large. For example, a horizontal (X) direction Sobel filter as shown in FIG. 10 performs an operation corresponding to a first-order differentiation in the horizontal (X) direction. When this Sobel filter is used, Where there is an edge in the (Y) direction, the output of the Sobel filter has a large positive or negative value, so that the edge in the vertical (Y) direction can be detected.
[0005]
In addition to the Sobel filter as described above, various spatial filters that perform operations corresponding to first-order or second-order differentiation are used for edge detection. As for these edge detection methods, for example, “Basics of Digital Image Processing” written by Jain, which is a typical textbook, and “Digital Image Processing” written by Pratt, ) ”.
[0006]
In addition, various methods for detecting the contour of the target object using the edge detection method as described above have been tried. For example, from the result of edge detection, a binary image is generated in which the degree of change in pixel value, that is, a portion where the edge strength is high is “1” and the other portion is “0”, and the binary image is thinned. A region extraction method for obtaining the contour of the target object by this is disclosed in Japanese Patent Application No. 5-233810. Here, the thinning means that the pixel of value “1” is sequentially removed from the generated binary image, and the pixel of value “1” is usually removed in order until the pixel width becomes 1 pixel width. Is a known procedure of extracting.
[0007]
On the other hand, there is also a method for detecting edges by a statistical method without using a spatial filter. For example, an edge detection method and apparatus that calculates a variance of pixel values in the vicinity of a pixel of interest, that is, a variance of hue, and detects a region where the value is large as a boundary of a region is disclosed in Japanese Patent Application No. 5-181969. Is disclosed.
[0008]
Also, the document “Color edge detection using vector orderistics” (Trahias, PE et al. “Color edge detection using vector order statistics”, IEEE Transactions on image processing, Vo1.2, Po1.2. 259-264, 1993) ”describes a method of detecting an edge based on a scale representing how far each pixel value is apart from other pixel values.
[0009]
[Problems to be solved by the invention]
However, the conventional edge detection method as described above has a problem of detecting all the portions where the pixel value has a sharp change. For example, as shown in FIG. 11A, in the target object 110 of the image 100, when there is a color difference or an intersection between the background and the target object 110 inside the target object 110, the conventional edge detection method uses the same figure. As shown in (b), the boundary 112 inside the object was also detected, and it was not possible to obtain only the edge 111 originally desired as shown in FIG.
[0010]
Usually, in an object, a sharp change in pixel value due to various reasons such as a color difference, a pattern, and a shadow as shown in FIG. 11A is unavoidable. Also, there are usually various patterns and colors in the background. On the other hand, since the conventional edge detection method uniformly detects steep changes in the pixel value, the above-mentioned boundary inside the object, the background pattern, etc. may be erroneously detected as edges. was there.
[0011]
Therefore, in order to solve the above-described problem, after performing weighting correction processing on the lightness based on the color information of the image, that is, lightness, saturation, and hue, the lightness subjected to the weighting correction processing is used. An automatic clipping system for detecting an edge based on this is disclosed in Japanese Patent Application No. 1-173177. This automatic cropping system uses a weighting correction process to emphasize changes in pixel values near the contour, for example, changes in brightness, thereby making the foreground object, background pattern and color change false. Edge detection is performed by reducing detection.
[0012]
However, in the automatic clipping system, it is not considered to perform an optimum weighting correction process for each individual portion on the contour. Also, it is rare that the pattern, color, etc. are uniform along the contour of the object. For this reason, even with the automatic clipping system, it is not possible to prevent the mixing of edges other than the edge that is originally desired.
[0013]
Further, an edge extraction device that detects an edge by highlighting an edge near the boundary by converting the pixel value so as to increase the contrast at the boundary between the foreground and the background using the histogram of the image is disclosed in Japanese Patent Application No. Hei 1 76170 ". However, although the above-described edge extraction apparatus is effective for detecting sensitive edges, it has not been possible to prevent the mixing of edges other than those originally desired.
[0014]
On the other hand, an area extracting apparatus that limits edge detection locations on an image, eliminates edge detection at unnecessary locations, and minimizes the inclusion of edges at locations other than contours is disclosed in “ 176780 ”. In this region extracting apparatus, an outline of a target object is manually input and edge detection is performed only within the outline outline, thereby preventing edge generation and performing edge detection. Also, a Japanese patent application No. 5-236347 discloses a soft key generation device configured to manually input an outline of an outline of a target object and perform edge detection only near the boundary based on the outline outline information. ing.
[0015]
However, the region extraction device and the soft key generation device simply narrow the region where erroneous edge detection occurs, and cannot prevent mixing of edges other than the contour.
For example, as illustrated in FIG. 12A, when edge detection is performed using an outline 220 of the target object 210 input by an operator in an image 200, the limit is limited even if the range of edge detection is limited. However, as shown in FIG. 12B, unnecessary edges 221a, 221b, and 221c are erroneously detected.
[0016]
Here, by increasing the accuracy of the outline outline 220, the number and length of unnecessary edges 221a, 221b, and 221c can be gradually reduced. For this purpose, the outline outline 220 is manually and highly accurate. This is a very difficult task. In order to solve this problem, when the operator slowly moves the input means such as a tablet or a mouse, the outline thickness of the outline is controlled to be thin. An image editing apparatus that detects edges by limiting the range of edge detection by controlling the outline to be thick is disclosed in Japanese Patent Application No. 1-180674. It was not possible to prevent the mixing of unnecessary edges.
[0017]
In order to solve the above problems, for example, the edge direction is estimated from a rough outline manually input, and only the edges that match the estimated edge direction are selectively detected, so that the above-described unnecessary There is an edge detection method in which edge detection is performed by removing a significant edge.
[0018]
In the edge detection method described above, only the edges in a specific direction are detected by using a filter group called Robinson's three-level (3-level) as shown in FIGS. .
The filter group shown in FIGS. 13A to 13D is an extension of the Sobel filter shown in FIG. 10 and detects edges in directions other than horizontal and vertical. For example, the filter shown in FIG. 13A outputs a large value when the pixel value on the right side of the image is large and the pixel value on the left side of the image is small. That is, the filter shown in FIG. 13A detects an edge in the + 0 ° direction. Further, the filters shown in FIGS. 13B, 13C, and 13D detect edges in the 45 °, 90 °, and 135 ° directions, respectively.
[0019]
In the edge detection method, in addition to the filter groups shown in FIGS. 13A to 13D, filters that detect edges in 45 ° increments in directions other than the edge directions detected by the filter groups. Groups are used. Therefore, the edge detection method uses a filter for detecting edges in eight directions in total.
[0020]
Here, in addition to the Robinson three-level filter shown in FIGS. 13A to 13D, as a filter for detecting an edge in a specific direction, “Digital Image Processing” by Pratt described above can be used. There are a Prewitt compass filter, a Kirsch filter, and the like described in “Digital Image Processing”.
[0021]
However, in the filters for detecting edges in a specific direction (hereinafter referred to as direction-selective filters) such as the above-described Robinson 3-level filter, Prewit compass filter, and Kirsch filter, the direction selectivity, That is, there is a drawback that directivity is weak.
[0022]
More specifically, for example, when the original intensity of the edge is K and the angle between the direction of the edge to be detected and the actual edge direction is θ, the output of the direction-selective filter, that is, the edge intensity E is
E = Kcosθ
Is given by the following relational expression.
[0023]
This relational expression indicates that the edge strength E of the unnecessary edge does not become “0” unless the direction of the unnecessary edge is 90 ° different from the direction of the edge to be originally detected, that is, the contamination of the unnecessary edge is completely removed. Indicates that you cannot. Therefore, as shown in FIG. 14, the directivity B that is practically necessary for the directivity A in the direction-selective filter needs to be stronger, but this strong directivity is the direction. Since the selective filter could not be provided, it was not possible to prevent the mixing of unnecessary edges.
[0024]
Further, in the direction-selective filter, when trying to make the edge detection direction fine, for example, when trying to make the edge detection direction fine in units of 15 ° instead of 45 ° as described above, the filter Was getting bigger. That is, the positive part and the negative part of the coefficient of the filter must be arranged at right angles to the edge detection direction. However, when the edge detection direction is finely specified in units of 15 °, the detection direction In order to represent this angle, a filter as small as 3 × 3 is insufficient, and the filter is large. For this reason, the ability to detect fine edges has been reduced, and inconveniences have arisen that it is impossible to deal with edges in all directions unless a large number of filters are used.
[0025]
Accordingly, the present invention has been made in view of the above-described conventional situation and has the following objects.
That is, an object of the present invention is to provide an edge detection method and an edge detection device that prevent the mixing of unnecessary edges and improve the edge detection accuracy.
[0026]
Another object of the present invention is to provide an edge detection method and an edge detection device that have strong directivity in the edge detection direction.
It is another object of the present invention to provide an edge detection method and an edge detection device that can easily obtain a good edge image.
[0027]
It is another object of the present invention to provide an edge detection method and an edge detection apparatus that can obtain a good edge image at high speed.
[0028]
[Means for Solving the Problems]
  In order to solve the above-described problem, the present invention is an edge detection method for detecting, as an edge, a pixel group whose pixel value changes sharply compared with the surroundings from input image data composed of grayscale data. There,By estimation meansA curve is generated along the point group to be subjected to edge extraction processing shown in advance, and for the points existing on the generated curve,In input image dataEstimate edge direction at each positionEstimating process and,In the input image data by the gradient calculation meansEach positionofSeeking a gradientGradient calculation process,By edge strength calculation means,the aboveIn the estimation processThe estimated edge direction and,The angle formed by the gradient, the size of the gradient,The smaller the absolute value of the angle between the estimated edge direction and the gradient, the larger the value.Index that determines directivityWhen,Edge detection by calculating edge strength fromThe edge strength calculation stepIt is characterized by that.
[0029]
  The edge detection method according to the present invention is characterized in that an edge strength is calculated by obtaining a cosine of an angle formed by the estimated edge direction and the gradient.
[0030]
  Further, the edge detection method according to the present invention normalizes the estimated edge direction and the length to “1”, which is a vector obtained by normalizing the cosine of the angle formed by the estimated edge direction and the gradient. It is obtained as an inner product with the gradient.
  Furthermore, the present invention is characterized in that the curve is a spline curve.
[0031]
  The present invention is an edge detection device for detecting, as an edge, a pixel group in which pixel values change sharply compared to the surroundings from input image data composed of grayscale data, the edge being shown in advance An estimation means for generating a curve along a point group to be extracted, and estimating an edge direction of each position in the input image data with respect to a point existing on the generated curve, and each position in the input image data Gradient calculation means for determining the gradient of the edge, the edge direction estimated by the estimation means, the angle formed by the gradient obtained by the gradient calculation means, the magnitude of the gradient, and the index that determines the directivity are used to determine the edge strength. Edge strength calculating means for calculating, and detecting means for detecting an edge from the edge strength obtained by the edge strength calculating means The provided, the index is characterized by taking as the absolute value of the angle of the estimated edge direction and the gradient is small, a large value.
[0032]
  Further, the edge detection apparatus according to the present invention is characterized in that the edge intensity calculation means calculates an edge intensity by obtaining a cosine of an angle formed by the estimated edge direction and the gradient.
[0033]
Further, in the edge detection apparatus according to the present invention, the edge strength calculating means calculates the estimated edge direction, which is a vector obtained by normalizing the cosine of the angle formed by the estimated edge direction and the gradient, and the length. It is obtained as an inner product with the gradient normalized to “1”.
[0034]
  Moreover, the edge detection apparatus according to the present invention isThe above curve is a spline curveIt is characterized by that.
[0035]
[Action]
  In the edge detection method according to the present invention,By estimation meansA curve is generated along the point group to be subjected to edge extraction processing shown in advance, and for the points existing on the generated curve,In input image dataEstimate edge direction at each positionShi,In the input image data by the gradient calculation meansEach positionofFind the gradient. AndBy edge strength calculation means,The estimated edge direction and,The angle formed by the gradient, the size of the gradient,The smaller the absolute value of the angle between the estimated edge direction and the gradient, the larger the value.Index that determines directivityWhen,Edge strength is calculated from the edge to detect the edge.
[0036]
  In the edge detection method according to the present invention, the estimated edge direction and the gradient are formed.Angle andOf the above gradientFrom the size and index that determines directivity, the edge strength iscalculate.
[0037]
In the edge detection method according to the present invention, the estimated edge direction and the cosine of the angle formed by the gradient are normalized to a preliminarily normalized vector, and the length is normalized to “1”. Obtained as inner product with the gradient.
[0038]
  In the edge detection apparatus according to the present invention, a curve is generated along a point group to be subjected to edge extraction processing indicated in advance, and each position in the input image data is detected with respect to a point existing on the generated curve. The edge direction is estimated by the estimation means, and the gradient of each position in the input image data is obtained by the gradient calculation means. Then, the edge strength calculation means calculates the edge strength from the edge direction estimated by the estimation means, the angle formed by the gradient obtained by the gradient calculation means, the magnitude of the gradient, and an index for determining directivity. Then, the edge is detected by the detecting means from the edge intensity obtained by the edge intensity calculating means. The index takes a larger value as the absolute value of the angle formed by the estimated edge direction and the gradient is smaller.
[0039]
  In the edge detection apparatus according to the present invention, the edge strength calculation means calculates the edge strength by obtaining a cosine of an angle formed by the estimated edge direction and the gradient.
[0040]
In the edge detection apparatus according to the present invention, the edge strength calculating means calculates the estimated edge direction and the length, which is a vector obtained by normalizing a cosine of an angle formed by the estimated edge direction and the gradient in advance. Obtained as an inner product with the gradient normalized to “1”.
[0042]
【Example】
Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
The edge detection method according to the present invention is implemented by, for example, an edge detection apparatus 10 as shown in FIG. The edge detection apparatus 10 is an application of the edge detection apparatus according to the present invention.
[0043]
That is, as shown in FIG._in, A spline curve generation unit 12 to which an outline P is supplied, a parameter coordinate generation unit 13 that generates an auxiliary variable t for the spline curve generation unit 12, and a gradient calculation unit 11 And an edge strength calculator 14 for calculating the edge strength E from the output from the spline curve generator 12. Further, the output from the spline curve generator 12 is also supplied to the gradient calculator 11.
[0044]
In addition, the gradient calculation unit 11 receives the input image F as shown in FIG._inAre provided with a filter computing unit 111 and a filter computing unit 112, respectively. The filter calculator 111 is a Sobel filter S in the X direction._XInput image F_inThe filter computing unit 112 is provided with a Sobel filter S in the Y direction._YInput image F_inIs to be applied. Each output from the filter calculator 111 and the filter calculator 112 is supplied to the edge strength calculator 14.
[0045]
The edge strength calculator 14 is supplied with a preset direction selection index s, and the edge strength calculator 14 outputs an output from the gradient calculator 11 as shown in FIG. Gradient table T to be referenced using_GA coordinate rotation unit 412 to which an output from the spline curve generation unit 12 is supplied, and a gradient table T_GThe inner product calculation unit 141 to which the result obtained in step S3 and the output from the coordinate rotation unit 412 are supplied, and the cosine table T to be referred to using the supplied direction selection index s and the output from the internal calculation unit 141._CAnd cosine table T_CAnd the gradient table T_GAnd a multiplier 143 for calculating the edge strength E from the result obtained in (1).
[0046]
Here, FIG. 4 is a flowchart showing an edge detection process in the edge detection apparatus 10. Hereinafter, it demonstrates concretely using the said FIGS. 1-3 and FIG.
The edge detection device 10 includes an input image F composed of grayscale data._inIs entered. On the other hand, the edge detection device 10 includes, for example, an input image F displayed on a display (not shown) by the operator._inBy using the tablet while watching, the outline P of the contour is input.
[0047]
The gradient calculation unit 11 receives the input image F_in(Step S1). The detailed description of the gradient calculation unit 11 will be described later.
On the other hand, the spline curve generation unit 12 sequentially captures the outline P, which is input by the operator as described above, as a coordinate sequence (x0, y0), (x1, y1), (x2, y2),. S2). Then, the spline curve generator 12 generates a smooth curve P (t) that passes through the coordinate sequences (x0, y0), (x1, y1), (x2, y2),. Step S3).
[0048]
More specifically, first, the curve P (t) is expressed in a cubic spline format. That is, the curve P (t) has an auxiliary variable t and a cubic polynomial x (t), y (t) of the auxiliary variable t,
P (t) = (x (t), y (t))
It is expressed by the expression When the auxiliary variable t changes from “0” to “1”, the trajectory curve P (t), that is, the spline curve P (t), is represented by a coordinate sequence (x0, y0), (x1, y1). , (X2, y2),... Are smoothly connected to make a round of the contour. Such spline curves are used in a wide range of fields including CAD.
[0049]
Therefore, the spline curve generation unit 12 has a cubic polynomial so that the spline curve P (t) smoothly connects the coordinate sequences (x0, y0), (x1, y1), (x2, y2),. Coordinate sequence (x0, y0), (x1, y1), (x2, y2),... (hereinafter referred to as coordinate sequence (xn, yn)) that sequentially incorporates the coefficients of x (t) and y (t). )). This coefficient determination method is, for example, the determination of the coefficient described in “Curves and surfaces for computer aided geometric design” by Farin, a representative document. Apply the method. This coefficient determination method is a method for determining a coefficient that passes through a point group when a point group is given, and is known.
[0050]
For example, the input image F_in5 is an image 250 as shown in FIG. 5A, that is, when a circular object 260 exists in the vicinity of the center of the image 250, the spline curve generator 12 causes the operator to move along the periphery of the circular object 260. Point group 271 on the outline indicated by_xyAs a coordinate sequence (xn, yn). Then, the spline curve generation unit 12 includes a point group 271._xyA curve 270 for interpolating the point group 271_xyA coefficient that passes through is determined, and a spline curve P (t) is generated.
[0051]
The spline curve P (t) generated by the spline curve generation unit 12 as described above is supplied to the gradient calculation unit 11.
Further, the spline curve generation unit 12 uses the generated spline curve P (t) as an outline of the outline, and performs the following processing on each pixel on the spline curve P (t) (step S4). At this time, the parameter coordinate generation unit 13 changes the auxiliary variable t of the curve P (t) from “0” to “1” with a small step size, and generates it for the spline curve generation unit 12.
[0052]
In the process of step S4 described below, each pixel on the spline curve P (t) is sequentially traced by the auxiliary variable t which is gradually changed from the parameter coordinate generation unit 13, and each pixel (x, y) It is a loop process repeated about.
First, the spline curve generation unit 12 substitutes the auxiliary variable t from the parameter coordinate generation unit 13 into the above-described expression “P (t) = (x (t), y (t))”, thereby obtaining the target pixel. X coordinate and Y coordinate (hereinafter referred to as target pixel (x, y)). Then, the spline curve generation unit 12 supplies the obtained target pixel (x, y) to the gradient calculation unit 11.
[0053]
Further, the spline curve generation unit 12 obtains the tangential direction of the spline curve P (t) in the obtained target pixel (x, y) as shown in FIG. This tangential direction is used as the approximate direction of the contour in the target pixel (x, y). That is, since the tangential direction is equal to the direction of the velocity vector V (t) obtained by differentiating the spline curve P (t), this velocity vector V (t) is defined as the tangential direction.
V (t) = d / dtP (t)
The following formula is used. Then, the length of the obtained velocity vector V (t) is normalized to “1”, and the tangent vector T is
T = V (t) / | V (t) |
(Step S4.1) and the obtained tangent vector T is supplied to the edge strength calculator 14.
[0054]
Next, the tangent vector T obtained by the spline curve generation unit 12 as described above is used as auxiliary information, and edge detection processing described below is performed using the auxiliary information (step S4.2).
First, the edge strength calculation unit 14 obtains a normal vector N obtained by rotating the tangent vector T from the spline curve generation unit 12 by 90 ° (step S4.2.1). This rotation has a positive direction counterclockwise with respect to the traveling direction of the trajectory of the spline curve P (t).
[0055]
That is, in the edge strength calculation unit 14, as shown in FIG. 3, the coordinate rotation unit 142 converts the tangent vector T from the spline curve generation unit 12 into a normal vector N rotated counterclockwise. The processing in the coordinate rotation unit 142 is very simple by simply multiplying the coefficients by 1 and -1.
[0056]
When the processing for obtaining the normal vector N as described above is expressed using an arithmetic expression, the normal vector N is expressed by using the matrix R shown in Equation 1.
[0057]
[Expression 1]

[0058]
N = RT
It becomes. Since the normal vector N obtained in this way is a normal vector of the spline curve P (t), if the direction of the edge is parallel to the estimated contour direction, the gradient described later is perpendicular to each. And the normal vector N of the spline curve P (t) are also parallel.
[0059]
Next, the gradient calculation unit 11 described above inputs the input image F captured in the process of step S1 for the 3 × 3 pixel region in the vicinity of the target pixel (x, y) from the spline curve generation unit 12._inThe pixel value I (x, y) is read, and an image gradient (hereinafter referred to as a gradient) G at the target pixel (x, y) is obtained.
[0060]
That is, in the gradient calculation unit 11, as shown in FIG._inSobel filter S in the X direction with respect to the pixel value I (x, y) of_XApply. The filter calculator 112 also reads the input image F that has been read._inSobel filter S in the Y direction with respect to the pixel value I (x, y) of_YApply. The gradient calculation unit 11 combines the outputs Dx and Dy of the filter calculator 111 and the filter calculator 112 to calculate the edge intensity as the gradient G (= (Dx, Dy)) of the image at the target pixel (x, y). It supplies to the part 14 (step S4.2.2).
[0061]
The gradient G obtained as described above is a vector indicating the direction of the maximum inclination when the shade of the image is regarded as the altitude, and is perpendicular to the direction of the edge. That is, the gradient G is G = grad I (x, y) = (Dx, Dy) with respect to the pixel value I (x, y) of the input image Fin.
It can be expressed by the following formula.
[0062]
Next, the edge strength calculation unit 14 obtains the length lenG of the gradient G obtained by the gradient calculation unit 11 and the normG normalized so that the length lenG of the gradient G becomes “1” (step S4). 2.3).
Here, as shown in FIG. 3, the edge strength calculation unit 14 performs the gradient table T._GIt has. This gradient table T_GStores the result of pre-calculating the length lenG of the gradient G and the normalized normG using the arithmetic expression shown in [Equation 2].
[0063]
[Expression 2]

[0064]
Specifically, the gradient table T_GIs a table in which the gradient G (= (Dx, Dy)) obtained by the gradient calculation unit 11 is input, and the length lenG of the gradient G and normG obtained by normalizing the gradient G are output. Further, the accuracy of the pixel value is normally 8 bits, and the same accuracy is used for the output Dx in the X direction and the output Dy in the Y direction._G16 bits are input together with the output Dx in the X direction and the output Dy in the Y direction. Therefore, the gradient table T for any combination of the output Dx in the X direction and the output Dy in the Y direction_G1 element is assigned. Therefore, the gradient table T_GIn each element, the value of the length lenG of the gradient G and the value of normG obtained by normalizing it are stored. Such a gradient table T_GBy using, a result can be easily obtained without performing a complicated calculation such as a square root. In addition, gradient table T_GSince the input to is 16 bits, the table size, that is, the number of elements is “65536”, which can be sufficiently applied to the current computer.
[0065]
Gradient table T as described above_GThus, the length lenG of the gradient G of the gradient G (= (Dx, Dy)) obtained by the gradient calculation unit 11 and the normG obtained by normalizing it are obtained. Then, normG obtained by normalizing the obtained length G of the gradient G is supplied to the inner product calculation unit 141, and the length lenG of the gradient G is supplied to the multiplier 143.
[0066]
Finally, the edge strength calculation unit 14 obtains the edge strength E (x, y) of the contour in the target pixel (x, y) as described above, the normal vector N, and the gradient table T._GWith a gradient G length lenG from, a normalized normG and a direction selection index s,
E (x, y) = lenG (normG · N)^S
It is calculated by the following formula.
[0067]
That is, in the edge strength calculation unit 14, as shown in FIG. 3, first, the inner product calculation unit 141 first calculates the inner product “normG · N” of the normal vector N and the normG obtained by normalizing the length lenG of the gradient G. Is obtained by multiplication. This result is the cosine “cos θ” of the angle θ formed by the gradient G and the normal vector N.
[0068]
Next, cosine table T_CUsing the supplied direction selection index s,
cos^Sθ
Ask for. This cosine table T_CIs a table that takes the cosine “cos θ” obtained by the inner product calculation unit 141 as an input and outputs the c-th power of cosine “cos θ” as an output, and a cosine table T_CCosine “cos θ” output from the S-th power (= cos^Sθ) is supplied to the multiplier 143. Here, in this embodiment, the direction selection index s is a fixed value set in advance, so that the cosine table T_CThe input to is the result of the inner product obtained by the inner product calculation unit 141, and the cosine table T_CIs a small table whose input is about 8 bits. Such a cosine table T_CBy using the cosine of the cosine “cos” without performing a complicated calculation^Sθ ”can be easily obtained.
[0069]
Finally, the multiplier 143 performs the gradient table T_GGradient G length lenG from Cosine table T_CCosine raised to S-th power “cos^Sis multiplied by θ, and the result of multiplication is output as edge strength E. Here, the direction selection index s supplied to the edge strength calculation unit 14 is an index that determines how much it selectively reacts only to an edge in a specific direction. If the direction selection index s is larger, the directivity becomes stronger correspondingly, so that it is possible to more strongly prevent the mixing of unnecessary edges, but the desired direction, that is, the estimated edge direction and the actual edge direction are If there is a slight deviation, edge strength is attenuated. Therefore, the direction selection index s is usually set to a value between 2 and 8.
[0070]
For example, as shown in FIG. 6, when the direction selection index s increases, the edge strength E (only when the gradient G and the normal vector N are in the same direction, that is, the value of the inner product “normG · N” is close to “1”. x, y) is a large value. That is, high directivity can be obtained. When the direction selection index s is set to “1”, the same directivity as that of the above-described Robinson 3-level filter, pre-wits compass filter, and Kirsch filter can be obtained.
[0071]
The processing in step S4 as described above is performed for each pixel (x, y) on the spline curve P (t) that gives an outline. Therefore, as shown in FIG._inEdge strength image F obtained from_outIs output (step S5).
As described above, in the present embodiment, in the process of obtaining the edge strength E, the cosine “cos θ” is obtained directly by calculating the inner product without directly obtaining the angle formed by the gradient G and the normal vector N. There is no need to perform complicated processing such as calculation of angular functions. In other words, the amount corresponding to the magnitude of the angle formed by the gradient G and the normal vector N can be calculated with a much smaller calculation amount of only two multiplications and one addition.
[0072]
Further, in the conventional edge detection method and edge detection apparatus, unnecessary mixing of the edges 221a to 221c due to the color difference in the foreground object and the intersection of the background objects shown in FIG. According to the present embodiment, the edge intensity in the direction deviated by 45 ° from the estimated edge direction can obtain 29% attenuation, and the edge intensity in the direction deviated by 60 ° can obtain only 50% attenuation. As shown in FIG. 6, when the direction selection index s is “4”, the edge strength in the direction shifted by 45 ° from the estimated edge direction is attenuated by 75%, and the strength of the edge in the direction shifted by 60 ° is 94% attenuation can be obtained. In addition, since the directions of the edges 221a to 221c that may be caused by the color change in the foreground and the background as described above are significantly different from the direction of the contour of the target object, the strong directivity as described above is used. Unnecessary edge contamination can be almost completely removed.
[0073]
Further, a tangent vector T that is an estimated value of the direction of the outline and an input image F_inSince the gradient G is obtained for each pixel on the contour, an edge along the contour can be selectively detected at each point on the contour.
For example, in the image 250 shown in FIG. 5A, the normal vector N is in a direction perpendicular to the curve 270 at locations A and B where there is a color difference in the circular object 260. On the horizontal line generated by the difference in color, the gradient G is perpendicular to the horizontal line. Therefore, the angle difference between the normal vector N and the gradient G is 90 °, and the edge strength E is “0” as shown in FIG. That is, the edge in the circular object 260 is not detected. On the other hand, since the edge representing the contour is along the curve 270, the gradient G at the locations A and B is in the horizontal direction. This is in the same direction as the normal vector N. Therefore, the edge representing the contour is strongly detected. Therefore, it is possible to detect only the edge originally desired without mixing unnecessary edges 221a, 221b, and 221c as shown in FIG.
[0074]
Further, the tangent vector T, which is information for estimating the edge direction, is given as auxiliary information to the edge strength calculation unit 14 so that the edge strength calculation unit 14 reacts strongly in a specific direction, thereby eliminating unnecessary edges. Only necessary edges can be detected with a high S / N ratio without mixing.
[0075]
Further, since the direction of the edge of the target object at each position can be calculated from the outline of the contour input by the operator, only the necessary edge can be detected without increasing the amount of work for the operator. .
In the above-described embodiment, the setting value of the direction selection index s is set to a value between 2 and 8, but the accuracy of the estimated direction of the edge is further increased and a larger value is set for the direction selection index s. Also good. Thereby, directivity can be further strengthened.
[0076]
Further, the value of the direction selection index s may be changed according to the situation by inputting the value of the direction selection index s from the outside. Usually, the value of the direction selection index s suffices if it can be selected in several ways._CThe number of additional bits may be about 4 bits. Further, the value of the direction selection index s may be limited to an integer, and the calculator may calculate the power as repetition of multiplication.
[0077]
In the embodiment described above, the edge strength is obtained using the power of the cosine “cos θ”, but the output of the inner product calculation unit 141 shown in FIG. 3 is IP, and the directivity distribution function is f. (Θ)
f (cos^-1IP)
And the result may be used to determine the edge strength. Thereby, for example, as shown in FIG. 8, when the angle θ between the estimated edge direction and the actual edge direction is equal to or smaller than the threshold value K, the edge strength is not attenuated and exceeds the threshold value K. In this case, directivity that sharply attenuates the edge strength can be obtained.
[0078]
More specifically, first, a desired distribution shape is defined, and the value on the horizontal axis of the graph shown in FIG. 8 is converted into cosine. Next, the decay rate for each cosine value, ie, “f (cos^-1IP) "is stored in the table. This table is a cosine table T used in the above-described embodiment._CUse instead of. Therefore, the edge strength can be obtained using the distribution function f (θ).
[0079]
Further, the distribution function f (θ) may have the direction selection index s as a variable. That is, the edge strength may be obtained using the distribution function f (θs). In this case, the direction selection index s is also input to the table, and the edge strength is obtained while changing the direction selection index s.
[0080]
In the above-described embodiment, the outline of the outline is the spline curve P (t) expressed in the cubic spline format. However, the outline may be expressed in the spline format of different orders. Also, a curve expressed in an implicit form format without using the auxiliary variable t
Q (x, y) = 0
It is good.
[0081]
Alternatively, the outline image drawn by the operator may be thinned, and the edge direction may be estimated from the positional relationship between adjacent pixels. In this case, in step S3 shown in FIG. 4, the outline image is thinned, and in step S4.1 shown in FIG. 4, a plurality of pixels adjacent to the target pixel in the thinned image are displayed. The direction of the contour at that position is obtained from the pixel array.
[0082]
Further, in step S4 shown in FIG. 4, a plurality of filters that selectively react to edges in a specific direction, for example, a compass operator, are used based on the estimated direction of the edges. The edge strength may be obtained by selecting a filter having the highest detection capability in the direction from the plurality of filters.
[0083]
In the above-described embodiment, information representing the outline outline is input by the operator. However, it is also possible to detect edges continuously from a plurality of temporally continuous images, that is, moving images. .
In this case, in the process of step S2 in the flowchart shown in FIG. 4, the outline is not captured as the coordinate sequence (xn, yn), but the immediately preceding image, the immediately preceding image, The contour in the current processing target image is estimated from the edge detection result obtained from the current image and the three pieces of information on the current processing target image.
[0084]
That is, as shown in FIG. 9, first, the current processing target image F_CImage F immediately before_PPixels (hereinafter referred to as marks) 31 at regular intervals on the edge 30 obtained as a result of the edge detection process for₁, 31₂, 31_Three, ... Next, each mark 31₁, 31₂, 31_Three,... Are defined as small blocks each including a mark. For example, mark 31₁Is marked 31₁A small block 32 is defined. The portion having the highest correlation with the data in the small block thus determined is the image F to be processed at present._CDetect from inside by block matching process.
[0085]
In this block matching process, first, for example, the pixel value in the small block 32 and the current image F to be processed are processed._CThe block 33 having the highest correlation is detected by comparing the pixel values in the blocks having the same size and determining the height of the correlation based on the sum of the differences. At this time, the pixel 34 in the block 33 is changed to the mark 31 of the small block 32.₁Is a pixel corresponding to. Mark 31₁Mark 31 other than₂, 31_Three,... In the same manner, the corresponding pixel is set as the current processing target image F._CDetect from inside.
[0086]
Therefore, in the process of step S3 of the flowchart shown in FIG. 4, the spline curve P (t) is generated using the series of corresponding pixel coordinate sequences obtained as described above.
As described above, the contour is continuously extracted from the current processing target image and the immediately preceding image by block matching processing, and edge detection is performed using the contour to automatically detect the moving image without operator intervention. The edge can be detected automatically. In this case, the block matching processing may be performed by increasing the number of images, not only the current processing target image and the previous image, but also the image used for the block matching processing. Thereby, since the calculation accuracy of the amount of movement of a block can be raised, an edge can be detected more correctly.
[0087]
Further, the key frame method may be applied to the process of step S2 in the flowchart shown in FIG. 4 for a plurality of images separated in time. That is, a method may be applied in which an operator gives an instruction of a shape for a plurality of images, and a computer substitutes an operator by interpolation for intermediate images. In this case, in the process of step S2, the operator sets a rough outline, and for an intermediate image group, a point group on the rough outline is interpolated to automatically generate a spline curve in the image. .
[0088]
As described above, by detecting edges in the current image to be processed based on a plurality of images and the result of past edge detection, the edge direction or color change can be more accurately performed using the correlation between frames. It is possible to obtain auxiliary information that is information. Therefore, the edge detection accuracy can be further improved, and the edge detection in the moving image can be performed without trouble.
[0089]
【The invention's effect】
  In the edge detection method according to the present invention,By estimation meansA curve is generated along the point group to be subjected to edge extraction processing shown in advance, and for the points existing on the generated curve,In input image dataEstimate edge direction at each positionShi,In the input image data by the gradient calculation meansEach positionofSeeking a gradient,By the gradient intensity calculation means,The estimated edge direction and,The angle formed by the gradient, the size of the gradient,The smaller the absolute value of the angle between the estimated edge direction and the gradient, the larger the value.Index that determines directivityWhen,Edge strength is calculated from the edge to detect the edge.
  Thereby, strong directivity can be obtained in edge detection, and only edges constituting the contour of the target object can be detected without mixing unnecessary edges. Therefore, the detection accuracy can be increased.
[0090]
In the edge detection method according to the present invention, the edge strength is calculated by obtaining the cosine of the angle formed by the estimated edge direction and the gradient. Thereby, the edge strength can be calculated with a small amount of calculation. Therefore, a good edge image can be easily obtained.
[0091]
  In the edge detection method according to the present invention, the edge strength is calculated from the angle between the estimated edge direction and the gradient, the magnitude of the gradient, and an index for determining the directivity. The intensity can be calculated. Therefore, a good edge image can be easily obtained.
[0092]
In the edge detection method according to the present invention, the estimated edge direction and the cosine of the angle formed by the gradient are normalized to a preliminarily normalized vector, and the length is normalized to “1”. Obtained as inner product with the gradient. Thereby, stronger directivity can be obtained, and the edge strength can be calculated with a smaller amount of calculation. Therefore, a good edge image can be obtained more easily.
[0094]
  In the edge detection apparatus according to the present invention, a curve is generated along a point group to be subjected to edge extraction processing indicated in advance, and each position in the input image data is detected with respect to a point existing on the generated curve. The edge direction is estimated by the estimation unit, and the gradient of each position in the input image data is obtained by the gradient calculation unit, the edge direction estimated by the estimation unit, the angle formed by the gradient obtained by the gradient calculation unit, and the above The edge strength is calculated by the edge strength calculation means from the gradient magnitude and the index for determining the directivity, and the edge is detected by the detection means from the edge strength obtained by the edge strength calculation means. The index takes a larger value as the absolute value of the angle formed by the estimated edge direction and the gradient is smaller.
  Thereby, strong directivity can be obtained in edge detection, and only edges constituting the contour of the target object can be detected without mixing unnecessary edges. Therefore, the edge detection accuracy can be increased.
[0095]
In the edge detection apparatus according to the present invention, the edge strength calculation means calculates the edge strength by obtaining a cosine of an angle formed by the estimated edge direction and the gradient. Thereby, the edge strength can be calculated with a small amount of calculation. Therefore, a good edge image can be easily obtained.
[0096]
  In the edge detection apparatus according to the present invention, the edge strength calculation means calculates the edge strength from the angle formed by the estimated edge direction and the gradient, the magnitude of the gradient, and an index for determining directivity. Therefore, the edge strength can be calculated with a small calculation amount. Therefore, a good edge image can be easily obtained.
[0097]
In the edge detection apparatus according to the present invention, the edge strength calculating means calculates the estimated edge direction and the length, which is a vector obtained by normalizing a cosine of an angle formed by the estimated edge direction and the gradient in advance. Obtained as an inner product with the gradient normalized to “1”. Thereby, stronger directivity can be obtained, and the edge strength can be calculated with a smaller amount of calculation. Therefore, a good edge image can be obtained more easily.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of an edge detection apparatus according to the present invention.
FIG. 2 is a block diagram illustrating a configuration of a gradient calculation unit of the edge detection apparatus.
FIG. 3 is a block diagram illustrating a configuration of an edge strength calculation unit of the edge detection device.
FIG. 4 is a flowchart showing edge detection processing in the edge detection apparatus.
FIG. 5 is a diagram for explaining a curve that gives a rough outline;
FIG. 6 is a graph showing a relationship between a direction selection index and edge strength.
FIG. 7 is a diagram for explaining a relationship between an input image and an edge strength image.
FIG. 8 is a diagram for explaining a case where edge strength is obtained using a directivity distribution function.
FIG. 9 is a diagram for explaining block matching processing;
FIG. 10 is a schematic diagram illustrating a Sobel filter for detecting an edge in a specific direction.
FIG. 11 is a diagram for explaining a result when edge detection is performed on an image having a color difference inside an object using a conventional edge detection method;
FIG. 12 is a diagram for explaining a result when edge detection is performed on an image having a color difference inside an object using a conventional edge detection method using an outline of an outline input by an operator.
FIG. 13 is a schematic diagram illustrating a Robinson 3-level filter group for detecting an edge in a specific direction.
FIG. 14 is a diagram for explaining a comparison between directivity in a conventional edge detection method and directivity necessary for practical use.
[Explanation of symbols]
10 Edge detection device
11 Gradient calculator
12 Spline curve generator
13 Parameter coordinate generator
14 Edge strength calculator

Claims (8)

An edge detection method for detecting, as an edge, a pixel group in which pixel values change sharply compared to surroundings from input image data composed of grayscale data,
The estimation means generates a curve along the point group to be subjected to edge extraction processing indicated in advance, and estimates the edge direction of each position in the input image data with respect to a point existing on the generated curve. An estimation process ;
The gradient calculation unit, a gradient calculation step asking you to gradient at each location in the input image data,
The edge strength calculating unit, the edge direction estimated by the estimation process, the angle of the gradient obtained in the gradient calculation step, the magnitude of the gradient, the absolute of the angle between the estimated edge direction and the gradient An edge detection method comprising: an index for determining directivity that takes a larger value as the value is smaller; and an edge strength calculation step of calculating an edge strength from the index and detecting an edge .   2. The edge detection method according to claim 1, wherein an edge intensity is calculated by obtaining a cosine of an angle formed by the estimated edge direction and the gradient.   The cosine of the estimated edge direction and the direction formed by the gradient is obtained as an inner product of the estimated edge direction which is a vector normalized in advance and the gradient whose length is normalized to “1”. The edge detection method according to claim 1.   The edge detection method according to claim 1, wherein the curve is a spline curve. An edge detection device that detects, as an edge, a pixel group in which pixel values change sharply compared to surroundings from input image data composed of grayscale data,
An estimation means for generating a curve along a point group to be subjected to edge detection processing shown in advance, and estimating an edge direction at each position in the input image data with respect to a point existing on the generated curve;
A gradient calculating means for obtaining a gradient of each position in the input image data;
Edge strength calculating means for calculating edge strength from an edge direction estimated by the estimating means, an angle formed by the gradient obtained by the gradient calculating means, a magnitude of the gradient, and an index for determining directivity;
Detecting means for detecting an edge from the edge strength obtained by the edge strength calculating means,
The edge detection apparatus according to claim 1, wherein the index takes a larger value as an absolute value of an angle formed by the estimated edge direction and the gradient is smaller.   6. The edge detection apparatus according to claim 5, wherein the edge strength calculating means calculates an edge strength by obtaining a cosine of an angle formed by the estimated edge direction and the gradient.   The edge intensity calculating means includes a vector obtained by normalizing a cosine of an angle formed by the estimated edge direction and the gradient in advance, and the estimated edge direction and the gradient having a length normalized to “1”. 6. The edge detection device according to claim 5, wherein the edge detection device is obtained as an inner product.   6. The edge detection apparatus according to claim 5, wherein the curve is a spline curve.

Images