JP4065978B2 - Image compression apparatus, image compression program, and image compression method - Google Patents

Description

このような変換係数を、量子化ステップ幅Ｑで量子化した場合、量子化後の変換係数がｋ番目の量子化値を取る確率ｐｋは、
【数２】

となる。このとき、量子化後の符号化レートＲ（Ｑ）は、理想的にはエントロピーに等しく、
【数３】

となる。
【００３８】
一方、量子化歪みＤ（Ｑ）は、２乗誤差を用いて、
【数４】

と表される。
ここで、ｉ番目のサブバンドの量子化ステップ幅をＱｉとし、そのときの符号化レートをＲｉ（Ｑｉ）とし、量子化歪みをＤｉ（Ｒｉ）と表す。
【００３９】
ここでは、サブバンドの符号化レートの総和を目標レートにするという拘束条件のもとで、量子化歪みによる視覚的な画質劣化を最小化する。そこで、ラグランジェの未定乗数λを用いて、次の関数Ｊを導入する。
【数５】

この関数Ｊの式中において、γｉは、サブバンドｉに属する変換係数の視覚重みに該当し、上述したステップＳ４で既に求めた値群γ3LL〜γ1HHである。
【００４０】
次に、この関数ＪをＲｉで偏微分して、関数Ｊの第１項が停留値となる条件式を求めると、
【数６】

となる。この式を変形することにより、
【数７】

が得られる。
この［７］式に適当なλの値を与えて解くことにより、サブバンド毎のＱｉがそれぞれ求まる。これらのＱｉに基づいて、
【数８】

を計算し、トータルの符号化レートＲλを求める。
【００４１】
この符号化レートＲλが目標符号化レートに一致するように、未定乗数λを微調整することにより、量子化ステップ幅Ｑｉが確定する。
（なお、上述した最適化の演算結果をデータテーブルとして予めまとめておいてもよい。この場合、量子化部２３は、このデータテーブルを、変数αおよび視覚重みに基づいて参照することにより、最適な量子化ステップ幅を即座に決定することができる。）
【００４２】
ステップＳ６： 量子化部２３は、このように最適化された量子化ステップ幅Ｑｉに従って、各サブバンドの変換係数を量子化する。
【００４３】
ところで、上述した量子化ステップ幅の最適化では、評価パラメータｋが大きいほど、低域サブバンドの量子化ステップ幅は狭く設定される。その結果、評価パラメータｋの大きな画像データほど、低域の圧縮歪み（ここでは量子化歪み）が小さくなる。
【００４４】
ステップＳ７： ビットモデリング部２４は、ＲＯＩ設定部２７に予め設定される選択領域のマスク画像を読み出す。ビットモデリング部２４は、この選択領域のマスク画像に基づいて、サブバンド毎に変換係数が選択領域に含まれるか否かを判定する。ビットモデリング部２４は、選択領域内に位置する量子化後の変換係数を、Ｓビット分だけシフトする。なお、このＳビットは、公知のマックスシフト法に従って定められるビット数である。
ステップＳ８： ビットモデリング部２４は、量子化後の変換係数をビットプレーンに分割する。ビットモデリング部２４は、最上位桁のビットプレーンから順に、ＪＰＥＧ２０００に規定される３通りの符号化パスを実行する。
ステップＳ９： 算術符号化部２５は、ビットモデリング部２４から符号化データを処理単位ごとに取得する。算術符号化部２５は、この符号化データに対して、２値算術符号であるＭＱコーダーを用いて、算術符号化を施す。
ステップＳ１０： ビットストリーム生成部２６は、符号化データをＳＮＲプログレッシブその他の優先順位に従って並べ替え、ビットストリームを生成する。
ステップＳ１１： ビットストリーム生成部２６は、このビットストリームを所定の符号量で打ち切り、圧縮ファイルとして出力する。記録部３２は、この圧縮ファイルをメモリカード３３に記録保存する。
【００４５】
上述した一連の動作により、電子カメラ１１の撮像動作および画像圧縮処理が完了する。
【００４６】
［第１の実施形態の効果など］
以上説明したように、電子カメラ１１は、ウェーブレット変換係数の標準偏差を用いて、タイル境界の目立ちやすさを予測する。このとき、タイル境界の目立ちやすい画像データほど、低域の圧縮歪みを低く抑えるように量子化ステップ幅を調整設定する。
【００４７】
その結果、タイル境界が視覚的に目立ちやすい画像データ（例えば青空などの絵柄）を適切に選別し、問題となる低域の圧縮歪みを予め抑えることが可能になる。
また逆に、タイル境界が視覚的に目立ちにくい画像データ（例えば細かく複雑な絵柄）を適切に選別し、適度に低域の圧縮歪みを許容することが可能になる。
【００４８】
次に、別の実施形態について説明する。
《第２の実施形態》
第２の実施形態の構成は、第１の実施形態（図１）と同一であるため、ここでの説明を省略する。
【００４９】
図４は、第２の実施形態における画像圧縮処理を示す図である。以下、図４に示すステップ番号に沿って、第２の実施形態の動作を説明する。
ステップＳ２１： マイクロプロセッサ３９は、圧縮対象の画像データを複数の分割領域（例えばタイル）に分けて抽出し、分割領域ごとに画素値（輝度、色、色差などの値）の標準偏差を求める。
ステップＳ２２： 抽出した分割領域全体に占める『標準偏差が所定の閾値以下となった分割領域』の割合を求め、その割合を画像データの平坦度Ｒとする。この平坦度Ｒが大きいほど、画像データの階調（色調も含む）の平坦さが高く、階調ズレが目立ちやすいと判断できる。
ステップＳ２３： この平坦度Ｒが大きいほど、画像伸張に際してタイル境界に生じる不連続な圧縮歪みが目立ちやすくなると予測される。そこで、マイクロプロセッサ３９は、平坦度Ｒの値が大きいほど、低域サブバンドの視覚重みを相対的に増大させる。マイクロプロセッサ３９は、このように調整設定された視覚重みを量子化部２３に伝達する。
ステップＳ２４： タイル処理部２１は、バッファメモリ１６内の画像データをタイル単位に分割して逐次読み出す。
ステップＳ２５： ウェーブレット変換部２２は、各タイルを独立にウェーブレット変換し、図３に示すようなサブバンドに分ける。
なお、以降の動作（ステップＳ２６〜３２）は、第１の実施形態で説明したステップＳ５〜Ｓ１１と同一動作のため、ここでの説明を省略する。
【００５０】
［第２の実施形態の効果など］
以上説明した動作により、第２の実施形態においても、第１の実施形態と同様の効果を得ることができる。
【００５１】
特に、第２の実施形態では、画素値の標準偏差に基づいて、タイル境界の目立ちやすさを予測している。その結果、第１の実施形態とは異なり、ウェーブレット変換を待たずに予測動作を実施することができる。その結果、タイル境界の目立ちやすさの予測動作と、ウェーブレット変換などを並列に処理できるなど、処理効率の向上に適している。
【００５２】
以下、別の実施形態について説明する。
《第３の実施形態》
第３の実施形態の構成は、第１の実施形態（図１）と同一であるため、ここでの説明を省略する。
【００５３】
図５は、第３の実施形態における画像圧縮処理を示す図である。以下、図５に示すステップ番号に沿って、第３の実施形態の動作を説明する。
ステップＳ４１〜Ｓ４２： 第１の実施形態で説明したステップＳ１〜Ｓ２と同じ。
ステップＳ４３： ウェーブレット変換部２２は、各タイルについて、ウェーブレット変換係数の分散をサブバンド毎に求める。下式は、ウェーブレット変換係数の分散（標準偏差の二乗）を算出するための式である。
【数９】

【００５４】
なお、ここではタイル内部の全域について分散を算出している。しかしながら、タイル境界と関係するタイル内周に限って分散を算出してもよい（つまり、タイル境界と関係しないタイルの中央域は分散演算から除いてもよい）。また、タイルの各辺ごとにタイル境界の目立ちやすさを区別して予測するため、タイルの上辺近傍、下辺近傍、左辺近傍、および右辺近傍に分けて、分散を算出してもよい。
【００５５】
図６は、テスト画像（人物画像）の特徴的な画像領域をピックアップして、ウェーブレット変換係数の標準偏差をサブバンド別に示した図である。
このテスト画像の場合、『背景』に該当するタイルは、画像全体がぼかされているため、高域サブバンドから低域サブバンドにかけて標準偏差の値が一様に低く、タイル境界が非常に目立ちやすいと判断できる。
【００５６】
一方、『ニットの服』，『目』，『口元』のタイルは、低域サブバンドから中域サブバンドにかけて標準偏差の値が比較的高く、タイル境界は比較的目立ちにくいと判断できる。
【００５７】
また、『髪の毛』のタイルは、低域サブバンドから高域サブバンドにかけて標準偏差の値が全体的に高い。そのため、複雑な絵柄に埋もれて、タイル境界は殆ど目立たないと判断できる。
【００５８】
ウェーブレット変換部２２は、このように求めた標準偏差を、マイクロプロセッサ３９に伝達する。
ステップＳ４４： マイクロプロセッサ３９は、図６で示したような評価判断を計算で行うため、タイル境界の目立ちやすさを数値で示す評価パラメータｋを求める。ここでの評価パラメータｋは、第１の実施形態とは異なり、下式を用いて、タイル別に算出された値である。
【数１０】

［１０］式中の重み係数Ｗｓは、各サブバンドＳがタイル境界の目立ちやすさに寄与する割合を示す値であり、画像の主観評価実験などに基づいて決定すればよい。図７は、このように決定した重み係数Ｗｓの一例を示す表である。また、［１０］式中のＡは比例定数である。
【００５９】
なお、ステップＳ４３において、タイルを上辺、下辺、右辺、および左辺に区分して分散を計算している場合は、これらの区分ごとに評価パラメータｋを求めることが好ましい。
【００６０】
ちなみに、ここでは［１０］式を用いて評価パラメータｋを決定しているが、本発明はこの評価式に限定されない。分散の値が大きいほど、タイル境界が目立たないとの評価パラメータを算出する評価式であればよい。
【００６１】
図８は、テスト画像（人物画像）の特徴的な画像領域をピックアップして、評価パラメータｋの値を示した表である。この評価パラメータｋの値が大きいほど、タイル境界が目立ちやすいと予測できる。
ステップＳ４５： 続いて、マイクロプロセッサ３９は、隣接タイル間において評価パラメータｋの値を調整する。ここでは、マイクロプロセッサ３９が、処理対象のタイルごとに上下左右の隣接タイルを含めた局所領域をまず設定する。次に、マイクロプロセッサ３９は、その局所領域内で評価パラメータｋの平滑化（単純平均や重み付き平均など）を行う。
【００６２】
図９［Ａ］は、人物画像の評価パラメータｋをタイル区分に合わせて配列したマップである。図９［Ｂ］は、平滑化処理後の評価パラメータｋを示したマップである。
なお、ここでの平滑化演算の代わりに、ＭＡＸ値，メディアン値などの算出演算を使用してもよい。
【００６３】
また、ステップＳ４３において、タイルを上辺、下辺、右辺、および左辺に区分して標準偏差を計算している場合は、実際に隣接する辺の間で評価パラメータｋの調整を行えばよい。（例えば、右に位置するタイルの左辺と、左に位置するタイルの右辺との間で、評価パラメータｋの調整を行えばよい。）
【００６４】
このような評価パラメータｋの調整により、タイル境界が目立ちやすいタイルまたは辺が存在した場合、そのタイルまたは辺に接するタイルについてもタイル境界が目立ちやすいとの判断がなされる。
ステップＳ４６： マイクロプロセッサ３９は、このような評価パラメータｋによる予測状況を加味し、次式を用いて各サブバンドの視覚重みを決定する。
γ3LL＝1.0
γ3LH＝γ3HL＝exp(-1/18×k)
γ3HH＝exp(-2/18×k)
γ2LH＝γ2HL＝exp(-4/18×k)
γ2HH＝exp(-8/18×k)
γ1LH＝γ1HL＝exp(-9/18×k)
γ1HH＝exp(-18/18×k)
上式では、評価パラメータｋが大きくなるに従って、低域サブバンドの視覚重みγ3LL が相対的に大きく設定される。
【００６５】
マイクロプロセッサ３９は、このように選択された視覚重みを、量子化部２３に伝達する。
ステップＳ４７： 量子化部２３は、タイル別に設定された視覚重みに従って、量子化歪みを視覚的な画質劣化を最小化するように、変換係数の量子化ステップをタイルごとに最適化する。この最適化手法の手順については、第１の実施形態のステップＳ５と同一であるため、ここでの重複説明を省略する。
ステップＳ４８： 量子化部２３は、タイル別に最適化された各サブバンドの量子化ステップ幅を用いて、タイル別に変換係数を量子化する。
【００６６】
ところで、上述した量子化ステップ幅の最適化では、評価パラメータｋが大きいほど、低域サブバンドの量子化ステップ幅は狭く設定される。その結果、評価パラメータｋの大きなタイルほど、低域の圧縮歪み（ここでは量子化歪み）が小さくなる。
ステップＳ４９〜Ｓ５３： 第１の実施形態で説明したステップＳ７〜Ｓ１１と同じ。
ステップＳ５４： マイクロプロセッサ３９は、画像伸張時のデータ使用に備えて、タイル別に決定した量子化ステップ幅に関するデータを圧縮データに付加し、圧縮ファイルを生成する。記録部３２は、生成された圧縮ファイルをメモリカード３３に記録保存する。
【００６７】
上述した一連の動作により、電子カメラ１１の撮像動作および画像圧縮処理が完了する。
【００６８】
［第３の実施形態の効果など］
以上説明した動作により、第３の実施形態においても、第１の実施形態と同様の効果を得ることができる。
【００６９】
特に、第３の実施形態では、タイル境界の目立ちやすいタイル（例えば青空などのタイル）を適切に選別し、問題となる低域の圧縮歪み（すなわちタイル境界の発生）をタイル単位に抑えることが可能になる。また逆に、タイル境界が視覚的に目立ちにくいタイル（例えば絵柄の複雑なタイル）を適切に選別し、適度に低域の圧縮歪みを許容することが可能になる。その結果、絵柄の複雑なタイルについては、タイル境界の発生を許容する分だけ、中域〜高域周波数成分の画像情報をなるべく正確に圧縮し、圧縮による画像情報劣化をできる限り防ぐことが可能になる。
【００７０】
このように第３の実施形態では、タイル境界の目立ちやすさを基準にして、圧縮符号量の周波数配分をタイル別に調整する。その結果、タイル境界が現実問題となる極めて高い圧縮率において、タイル境界を改善しつつ、画像情報劣化を極力抑えるといった調整を実現することが可能になる。
【００７１】
さらに、第３の実施形態では、タイル境界の目立つタイルの周囲についても、低域の圧縮歪みを抑えている。その結果、平坦タイルの周囲に低域歪みが発生して、結果としてタイル境界が目立ってしまうといった事態を改善することが可能になる。
【００７２】
なお、上述した実施形態では、変換係数や画素値の標準偏差（分散の平方根）を求めて、階調ズレの目立ちやすさを評価している。しかしながら、本発明はこれに限定されるものではない。一般に、本発明の統計的性質としては、『階調ズレの目立ちやすさ』と相関する量であればなんでもよい。例えば、階調の平坦さ、階調の一様さ、階調の単調さなどを示す統計的性質でもよい。
【００７３】
また、本発明の『分散』は、値の分散具合を示す尺度であればよく、［９］式で示した統計学的な算出式に限定されるものではない。
【００７４】
ちなみに、画素値のダイナミックレンジは、瞬間的にでも大きな階調変化があれば大きく変動してしまう値である。そのため、画素値のダイナミックレンジについては、階調の平坦さを示す値ではなく、『階調ズレの目立ちやすさ』と相関する量でもない。
【００７５】
また、上述した実施形態では、画像圧縮パラメータとして、変換係数の量子化ステップ幅を調整設定している。しかしながら、本発明はこれに限定されるものではない。一般に、本発明の画像圧縮パラメータは、タイル境界における圧縮歪みの出現量を左右するパラメータであればなんでもよい。例えば、画像圧縮パラメータとしては、ビットプレーンの下位桁打ち切りを調整するパラメータや、ビットストリームの打ち切りを調整するパラメータや、ＲＯＩ符号化の選択領域を調整するパラメータなどでもよい。
【００７６】
なお、第１および第２の実施形態では、複数のタイル全体に対して画像圧縮パラメータを一律に変更している。この場合、複数のタイル全体に均質に画像圧縮を施すため、画面上において圧縮歪みの現れ方にムラを生じるなどのおそれが少ない。ただし、本発明はこれに限定されるものではない。例えば、第３の実施形態に示したように、個々のタイルごとに画像圧縮パラメータを調整設定してもよい。また、タイル境界とタイル内側とにおいて画像圧縮パラメータを個別に調整設定してもよい。
【００７７】
また、上述した実施形態では、電子カメラ１１において本発明の画像圧縮を実施する場合について説明した。しかしながら、本発明はこれに限定されるものではない。例えば、コンピュータ上において本発明の画像圧縮装置を実現してもよい。この場合、上述した実施形態における画像圧縮の手順をプログラムコード化することにより、画像圧縮プログラムを作成すればよい。
【００７８】
なお、本発明は、その精神または主要な特徴から逸脱することなく、他のいろいろな形で実施することができる。そのため、前述の実施例はあらゆる点で単なる例示に過ぎず、限定的に解釈してはならない。本発明の範囲は、特許請求の範囲によって示すものであって、明細書本文には、なんら拘束されない。さらに、特許請求の範囲の均等範囲に属する変形や変更は、すべて本発明の範囲内のものである。
【００７９】
【産業上の利用の可能性】
本発明の画像圧縮では、「階調ズレの視覚的な目立ちやすさ」と相関する統計的性質を画像データから求め、この統計的性質により予測されるタイル境界の目立ちやすさに応じて、圧縮歪みの出現量を左右する画像圧縮パラメータを調整設定する。
したがって、画像伸張時にタイル境界が目立ちそうな画像データに対しては、圧縮歪みを予め抑えるなどの適切な対処が容易になる。
【００８０】
【図面の簡単な説明】
【図１】本実施形態における電子カメラ１１の構成を示す図である。
【図２】第１の実施形態における画像圧縮処理を説明する流れ図である。
【図３】ウェーブレット変換によるサブバンド分割の様子を示す図である。
【図４】第２の実施形態における画像圧縮処理を説明する流れ図である。
【図５】第３の実施形態における画像圧縮処理を説明する流れ図である。
【図６】テスト画像の特徴的なタイルをピックアップして、変換係数の標準偏差（分散の平方根）をサブバンド別に示した図である。
【図７】評価パラメータｋの計算に使用するサブバンド別の重み係数Ｗｓの一例を示す図である。
【図８】テスト画像の特徴的なタイルをピックアップして、評価パラメータｋを示した図である。
【図９】評価パラメータｋをタイル区分に従って配列したマップである。[0001]
【Technical field】
  The present invention relates to an image compression apparatus that divides image data into a plurality of tiles and compresses the image for each tile.. BookThe present invention relates to an image compression program for causing a computer to function as an image compression apparatus.. BookThe present invention relates to an image compression method for compressing image data for each tile.
[0002]
[Background]
  Conventionally, as in JPEG 2000, an image compression method is known in which image data is divided into small areas called tiles and image compression is performed for each tile.
  Such tile-based image compression has the advantage that it is not necessary to handle a large amount of image data at a time, thus saving work memory.
[0003]
  Further, even if the screen size (number of vertical and horizontal pixels) of the image data varies, the screen size of each tile can be kept substantially constant by changing the number of tile divisions. In this case, there is an advantage that various screen sizes can be flexibly handled without changing the image compression processing in units of tiles.
  Furthermore, since the image compression processing can be separately executed for each tile, there is an advantage that parallel processing of image compression, multithreading, and the like can be easily realized.
[0004]
  Due to the advantages as described above, image compression in units of tiles is an important technology in image compression technology.
[0005]
  Although different from tile-based image compression, the following Patent Document 1 is known as a conventional technique for compressing an image in pixel block units, which are orthogonal transform division ranges. In Patent Document 1, an activity indicating the degree to which a spatial frequency component is included is calculated for each pixel block, and the compression code amount assigned to each pixel block is adjusted according to the activity.
[Patent Document 1]
JP-A-2-262786 (US5051840)
[0006]
DISCLOSURE OF THE INVENTION
[Problems to be solved by the invention]
  By the way, in the above-described tile-unit image compression, image compression is performed individually on adjacent tiles. Therefore, compressive distortion is also generated independently for each tile, and the compressive distortion is discontinuous at the tile boundary. As a result, the expanded image has a problem that a boundary between tiles appears as a boundary (hereinafter, this boundary is referred to as a “tile boundary”).
[0007]
  In order to improve such a tile boundary problem, image compression close to lossless compression may be performed to suppress the occurrence of compression distortion itself. However, if the compression distortion is suppressed more than necessary, the encoding efficiency of image compression becomes low, and problems such as a large increase in the amount of compressed data occur.
[0008]
  Further, in Patent Document 1 described above, the amount of compression code to be allocated is increased as the pixel block has a higher activity. Therefore, the amount of compression code to be allocated is reduced in the flat portion where the tile boundary is most conspicuous (the activity is almost zero). As a result, in Patent Document 1, it is difficult to positively improve the tile boundary.
[0009]
  Furthermore, in Patent Document 1, only the compression code amount is adjusted in units of pixel blocks, and the frequency distribution of the compression code amount (particularly, the code amount allocated to the low frequency component) is not changed. For this reason, it seems that it was difficult to make fine adjustments such as improving the tile boundary and suppressing image information deterioration in the middle and high frequency range at an extremely high image compression rate where the tile boundary becomes a real problem.
[0010]
  In view of the above-described problems, an object of the present invention is to provide an image compression technique capable of appropriately suppressing tile boundaries so that they are not conspicuous during image expansion.
[0011]
[Means for Solving the Problems]
  The present invention has been invented to solve the above-described problems, and the image compression apparatus according to the first aspect of the present invention divides image data into a plurality of tiles and performs image compression for performing image compression for each tile. A statistical processing unit that processes the image data and obtains a statistical property that correlates with “ease of conspicuous gradation shift in the case where gradation shift occurs in the image data”; A parameter setting unit that predicts the conspicuousness of the tile boundary at the time of image expansion based on the physical properties, and adjusts and sets the “image compression parameter that affects the appearance amount of compression distortion” according to the conspicuousness of the tile boundary; An image compression unit that compresses the image data using the adjusted image compression parameter, and the statistical property is a conversion coefficient obtained by wavelet transforming the image data. A stomach obtained dispersion,The image compression parameter is a quantization step width;The parameter setting unit calculates the wavelet transform coefficient.For each subbandDispersion with higher weighting as it goes around the tile reduces the compression distortion of the low-frequency spatial frequency component as it decreases.Narrow the quantization step width of the low frequency subbandIt is characterized by setting.
[0012]
  The second invention is the first invention, whereinThe statistical processing unit obtains the statistical property for each tile unit, and the parameter setting unit predicts the conspicuousness of the tile boundary for each tile unit based on the statistical property of the tile unit, Adjust the image compression parameter in a direction that reduces the compression distortion of the low-frequency spatial frequency component for tiles that are more conspicuous at the tile boundary.It is characterized by.
[0013]
  Further, the third invention is the first2In the invention ofThe parameter setting unit adjusts and sets the image compression parameter in a direction to reduce compression distortion of a low-frequency spatial frequency component even for a tile adjacent to the tile in which the tile boundary is conspicuous.It is characterized by.
[0014]
  The fourth inventionThe image compression program causes a computer to function as the statistical processing unit, the parameter setting unit, and the image compression unit according to any one of the first to third inventions.
[0015]
  The fifth inventionThe image compression method is an image compression method in which image data is divided into a plurality of tiles and image compression is performed for each tile, and the image data is processed to generate “a gradation shift in the image data. A statistical processing step for obtaining a statistical property correlating with the `` easy to notice the gradation shift in the case of the image, '' predicting the conspicuousness of the tile boundary at the time of image expansion based on the statistical property, A parameter setting step for adjusting and setting “an image compression parameter that affects the appearance amount of compression distortion” according to conspicuousness, and an image compression step for compressing the image data using the adjusted image compression parameter And the statistical property is a variance obtained with respect to a transform coefficient obtained by wavelet transforming the image data, and the image compression parameter is a quantity. The parameter setting unit is configured so that the dispersion of the weighting increases as the wavelet transform coefficient is moved around the tile for each subband, so that the smaller the variance, the smaller the compression distortion of the low-frequency spatial frequency component. The quantization step width of the low-frequency subband is set to be narrow.
[0017]
【The invention's effect】
  According to the present invention, when image data whose tile boundaries are visually conspicuous is appropriately selected, it is possible to suppress low-frequency compression distortion in advance. On the other hand, when image data whose tile boundaries are visually inconspicuous is appropriately selected, it is possible to appropriately allow low-frequency compression distortion.
[0018]
DETAILED DESCRIPTION OF THE INVENTION
  Embodiments according to the present invention will be described below with reference to the drawings.
[0019]
<< First Embodiment >>
[Description of electronic camera configuration]
  FIG. 1 is a diagram showing a configuration of an electronic camera 11 in the present embodiment.
  In FIG. 1, a photographing lens 12 is attached to the electronic camera 11. In the image space of the photographic lens 12, the light receiving surface of the image sensor 13 is arranged. The output of the image sensor 13 is processed through the AD converter 14 and the signal processor 15 and then recorded in the buffer memory 16.
  An image processing unit 17 and an image compression unit 18 are respectively connected to the data bus of the buffer memory 16.
[0020]
  This image compression unit 18 has the following configuration requirements:1) ~7)Is provided.
1)Tile processing unit 21
2)Wavelet transform unit 22
3)Quantization unit 23
4)Bit modeling unit 24
5)Arithmetic coding unit 25
6)Bitstream generation unit 26
7)ROI setting unit 27
[0021]
  The compressed file generated in the image compression unit 18 is given to the recording unit 32. The recording unit 32 records and saves this compressed file in the memory card 33.
[0022]
  Further, an operation member 38 including a release button and a setting button is provided on the casing of the electronic camera 11. Output signals of these operation members 38 are given to the microprocessor 39. The microprocessor 39 is a microprocessor for system control of the electronic camera 11, and is connected to the image compression unit 18, the image processing unit 17, and the like via signal lines..
[0023]
[Description of Operation of First Embodiment]
  The operation of the first embodiment will be described below.
  First, analog image data generated by the image sensor 13 is digitized via the AD conversion unit 14 and then supplied to the signal processing unit 15. The signal processing unit 15 performs signal processing such as black level correction and gamma correction on the image data. The image data after the signal processing is temporarily recorded in the buffer memory 16. The image processing unit 17 reads the image data in the buffer memory 16 for each processing unit, and performs image processing such as color interpolation processing and color coordinate conversion.
[0024]
  The image compression unit 18 reads out image data after image processing from the buffer memory 16 and performs image compression processing. FIG. 2 is a flowchart for explaining this image compression processing.
[0025]
  The image compression processing that is a feature of the present invention will be described below in the order of step numbers shown in FIG.
Step S1: The tile processing unit 21 in the image compression unit 18 sequentially reads out the image data in the buffer memory 16 by dividing it into tile units.
Step S2: The wavelet transform unit 22 performs wavelet transform on each tile and divides it into a plurality of subbands as shown in FIG.
Step S3: The wavelet transform unit 22 obtains the standard deviation (square root of variance) of the wavelet transform coefficient for each subband for each tile. Generally, the smaller the standard deviation is, the higher the flatness of the gradation (including the color tone) of the image data. For this reason, even if a slight gradation shift occurs in the image data, it can be determined that the gradation shift is conspicuous. The wavelet transform unit 22 transmits the standard deviation thus obtained to the microprocessor 39.
Step S4: The microprocessor 39 predicts the conspicuousness of the discontinuous compression distortion generated at the tile boundary during image expansion according to the standard deviation of the transform coefficient. The microprocessor 39 adjusts and sets the visual weight (visual importance) of each subband according to the prediction situation.
[0026]
  A specific example of this adjustment setting will be described below.
  First, the microprocessor 39 obtains a ratio of gradation flat tiles whose standard deviation is equal to or less than a threshold (here, 3 or less) in the 3LL band (lowest band) of each tile, and sets the ratio to r1. As the value of r1 is larger, the gradation of the image data is generally flat, and it can be determined that gradation deviation is more conspicuous.
[0027]
  Further, the microprocessor 39 obtains the ratio of the flat gradation tiles whose standard deviation is less than or equal to the threshold value (here, 3 or less) in the 3LH, 3HL, and 3HH bands of each tile, and sets the ratio to r2. This r2 is the ratio of tiles that can be regarded as flat gradation in the intermediate spatial frequency range. It can be determined that the larger the value of r2 in addition to r1, the flatter the gradation of the image data, and the more noticeable gradation deviation.
[0028]
  In addition, the microprocessor 39 obtains a ratio of gradation flat tiles whose standard deviation is equal to or less than a threshold (here, 3 or less) in the 2LH, 2HL, and 2HH bands of each tile, and sets the ratio to r3. This r3 is also the ratio of tiles that can be regarded as flat gradation in the intermediate spatial frequency range. In addition to r1 and r2, it can be determined that the larger the value of r3, the flatter the gradation of the image data and the more noticeable gradation shift.
[0029]
  Further, the microprocessor 39 obtains the ratio of the flat gradation tiles whose standard deviation is equal to or less than the threshold value (here, 3 or less) in the 1LH, 1HL, and 1HH bands of each tile, and sets the ratio to r4. This r4 is the ratio of tiles that can be regarded as flat gradation in the high spatial frequency range. In addition to r1 to r3, it can be determined that as the value of r4 is larger, the gradation of the image data is flat like silk, and the gradation shift is very conspicuous.
[0030]
  Using the gray level flat tile ratios r1 to 4 obtained in this way, the microprocessor 39
Parameter k = 8 (2 × r1 + 0.5 × r2 + 0.25 × r3 + 0.25 × r4)
And k is obtained. The parameter k is an evaluation parameter obtained by integrating the above-described r1 to r4. Therefore, it can be determined that as the evaluation parameter k is larger, the gradation of the image data is flat from the low frequency to the high frequency, and the gradation deviation is more conspicuous.
[0031]
  As such image data having a large evaluation parameter k, for example, a pattern such as a blue sky in a fine sky is assumed. Since such image data has a monotonous pattern, it is predicted that the tile boundary will rise and it will be particularly noticeable.
[0032]
  On the other hand, as the image data having a small evaluation parameter k, a complicated pattern including abundant details is assumed. In such image data, it is predicted that the discontinuous compression distortion generated at the tile boundary is buried in a complicated pattern and is not noticeable.
[0033]
  The microprocessor 39 considers such a prediction situation based on the evaluation parameter k, and determines the visual weight of each subband using the following equation.
γ3LL = 1.0
γ3LH = γ3HL = exp (-1 / 18 × k)
γ3HH ＝ exp (-2 / 18 × k)
γ2LH = γ2HL = exp (-4 / 18 × k)
γ2HH ＝ exp (-8 / 18 × k)
γ1LH ＝ γ1HL ＝ exp (-9 / 18 × k)
γ1HH ＝ exp (-18 / 18 × k)
[0034]
  In the above equation, the visual weight γ3LL of the low frequency sub-band is set relatively large as the evaluation parameter k increases.
[0035]
  This is a treatment based on the knowledge that the inventor has made clear based on a number of image compression experiments that “the lower the compression distortion occurs, the more easily the tile boundary becomes noticeable”.
[0036]
  The microprocessor 39 transmits the visual weight thus selected to the quantization unit 23.
Step S5: The quantizing unit 23 optimizes the quantization step width of the transform coefficient so as to minimize visual image quality degradation due to quantization distortion under an appropriate encoding rate constraint.
[0037]
  In the following, the principle of optimizing the quantization step width will be described in principle.
  Usually, the transform coefficient for each subband shows a probability distribution that is very similar to the Laplace distribution. Therefore, the probability distribution of the transform coefficient is approximated by the following Laplace distribution formula to determine a variable α that identifies the Laplace distribution.
[Expression 1]

  When such a transform coefficient is quantized with the quantization step width Q, the probability pk that the transform coefficient after quantization takes the kth quantized value is:
[Expression 2]

It becomes. At this time, the coding rate R (Q) after quantization is ideally equal to entropy,
[Equation 3]

It becomes.
[0038]
  On the other hand, the quantization distortion D (Q) uses the square error,
[Expression 4]

It is expressed.
  Here, the quantization step width of the i-th subband is Qi, the coding rate at that time is Ri (Qi), and the quantization distortion is represented by Di (Ri).
[0039]
  Here, visual image quality degradation due to quantization distortion is minimized under the constraint that the sum of the subband coding rates is the target rate. Therefore, the following function J is introduced using Lagrange's undetermined multiplier λ.
[Equation 5]

  In the expression of the function J, γi corresponds to the visual weight of the transform coefficient belonging to the subband i, and is the value group γ3LL to γ1HH already obtained in step S4 described above.
[0040]
  Next, when this function J is partially differentiated by Ri, a conditional expression in which the first term of the function J is a stationary value is obtained.
[Formula 6]

It becomes. By transforming this equation,
[Expression 7]

Is obtained.
  Qi for each subband can be obtained by solving the equation [7] by giving an appropriate value of λ. Based on these Qi,
[Equation 8]

To calculate the total coding rate Rλ.
[0041]
  The quantization step width Qi is determined by finely adjusting the undetermined multiplier λ so that the coding rate Rλ matches the target coding rate.
(Note that the optimization calculation results described above may be preliminarily compiled as a data table. In this case, the quantization unit 23 refers to the data table based on the variable α and the visual weight to optimize the data table. The correct quantization step width can be determined immediately.)
[0042]
Step S6: The quantization unit 23 quantizes the transform coefficient of each subband according to the quantization step width Qi optimized in this way.
[0043]
  By the way, in the above-described optimization of the quantization step width, the quantization step width of the low frequency subband is set narrower as the evaluation parameter k is larger. As a result, the lower the compression distortion (quantization distortion in this case) of the image data having a larger evaluation parameter k.
[0044]
Step S <b> 7: The bit modeling unit 24 reads a mask image of a selection area that is preset in the ROI setting unit 27. The bit modeling unit 24 determines whether a conversion coefficient is included in the selection region for each subband based on the mask image of the selection region. The bit modeling unit 24 shifts the quantized transform coefficient located in the selected region by S bits. The S bit is the number of bits determined according to a known max shift method.
Step S8: The bit modeling unit 24 divides the quantized transform coefficient into bit planes. The bit modeling unit 24 executes three encoding passes defined in JPEG2000 in order from the most significant bit plane.
Step S9: The arithmetic encoding unit 25 acquires encoded data from the bit modeling unit 24 for each processing unit. The arithmetic encoding unit 25 performs arithmetic encoding on the encoded data using an MQ coder that is a binary arithmetic code.
Step S10: The bit stream generation unit 26 rearranges the encoded data according to SNR progressive or other priorities to generate a bit stream.
Step S11: The bit stream generation unit 26 terminates this bit stream with a predetermined code amount and outputs it as a compressed file. The recording unit 32 records and saves this compressed file in the memory card 33.
[0045]
  The imaging operation and image compression processing of the electronic camera 11 are completed by the series of operations described above.
[0046]
[Effects of First Embodiment]
  As described above, the electronic camera 11 predicts the conspicuousness of the tile boundary using the standard deviation of the wavelet transform coefficient. At this time, the quantization step width is adjusted and set so that the image data that is more conspicuous at the tile boundary is lower in low-frequency compression distortion.
[0047]
  As a result, it is possible to appropriately select image data (for example, a pattern such as a blue sky) in which tile boundaries are visually conspicuous, and to suppress a low-frequency compression distortion in advance.
  On the other hand, it is possible to appropriately select image data (for example, a fine and complicated pattern) in which tile boundaries are not visually noticeable, and to allow moderately low-frequency compression distortion.
[0048]
  Next, another embodiment will be described.
<< Second Embodiment >>
  Since the configuration of the second embodiment is the same as that of the first embodiment (FIG. 1), description thereof is omitted here.
[0049]
  FIG. 4 is a diagram illustrating image compression processing according to the second embodiment. The operation of the second embodiment will be described below along the step numbers shown in FIG.
Step S21: The microprocessor 39 extracts the image data to be compressed by dividing it into a plurality of divided areas (for example, tiles), and obtains a standard deviation of pixel values (values such as luminance, color, and color difference) for each divided area.
Step S22: A ratio of “the divided area where the standard deviation is equal to or less than a predetermined threshold” in the extracted divided areas is obtained, and the ratio is set as the flatness R of the image data. It can be determined that the higher the flatness R, the higher the flatness of the gradation (including the color tone) of the image data, and the more noticeable gradation shift is.
Step S23: It is predicted that the greater the flatness R, the more conspicuous is the discontinuous compression distortion that occurs at the tile boundary during image expansion. Therefore, the microprocessor 39 relatively increases the visual weight of the low frequency sub-band as the flatness value R increases. The microprocessor 39 transmits the visual weight adjusted and set in this way to the quantization unit 23.
Step S24: The tile processing unit 21 sequentially reads out the image data in the buffer memory 16 divided into tiles.
Step S25: The wavelet transform unit 22 performs wavelet transform on each tile independently and divides it into subbands as shown in FIG.
  The subsequent operations (steps S26 to S32) are the same as steps S5 to S11 described in the first embodiment, and thus description thereof is omitted here.
[0050]
[Effects of Second Embodiment, etc.]
  With the operation described above, the same effects as in the first embodiment can be obtained in the second embodiment.
[0051]
  In particular, in the second embodiment, the conspicuousness of tile boundaries is predicted based on the standard deviation of pixel values. As a result, unlike the first embodiment, the prediction operation can be performed without waiting for the wavelet transform. As a result, it is suitable for improving the processing efficiency, such as predicting the conspicuousness of the tile boundary and processing the wavelet transform in parallel.
[0052]
  Hereinafter, another embodiment will be described.
<< Third Embodiment >>
  Since the configuration of the third embodiment is the same as that of the first embodiment (FIG. 1), description thereof is omitted here.
[0053]
  FIG. 5 is a diagram illustrating image compression processing according to the third embodiment. The operation of the third embodiment will be described below along the step numbers shown in FIG.
Steps S41 to S42: The same as steps S1 to S2 described in the first embodiment.
Step S43: The wavelet transform unit 22 obtains the variance of the wavelet transform coefficient for each subband for each tile. The following expression is an expression for calculating the variance (square of the standard deviation) of the wavelet transform coefficient.
[Equation 9]

[0054]
  Here, the variance is calculated for the entire area inside the tile. However, the variance may be calculated only for the inner circumference of the tile related to the tile boundary (that is, the central area of the tile not related to the tile boundary may be excluded from the variance calculation). In addition, in order to distinguish and predict the conspicuousness of the tile boundary for each side of the tile, the variance may be calculated separately for the vicinity of the upper side, the lower side, the left side, and the right side of the tile.
[0055]
  FIG. 6 is a diagram showing the standard deviation of the wavelet transform coefficient for each subband by picking up a characteristic image region of the test image (person image).
  In the case of this test image, the tile corresponding to the “background” has the entire image blurred, so the standard deviation value is uniformly low from the high frequency subband to the low frequency subband, and the tile boundary is very low. It can be judged that it is conspicuous.
[0056]
  On the other hand, the tiles of “knit clothes”, “eyes”, and “mouth” have a relatively high standard deviation value from the low-frequency subband to the mid-frequency subband, and it can be determined that the tile boundary is relatively inconspicuous.
[0057]
  In addition, the “hair” tile has a high standard deviation value from the low frequency subband to the high frequency subband. Therefore, it can be determined that the tile boundary is hardly noticeable because it is buried in a complicated pattern.
[0058]
  The wavelet transform unit 22 transmits the standard deviation thus obtained to the microprocessor 39.
Step S44: The microprocessor 39 obtains an evaluation parameter k indicating numerically the conspicuousness of the tile boundary in order to make the evaluation judgment as shown in FIG. 6 by calculation. Unlike the first embodiment, the evaluation parameter k here is a value calculated for each tile using the following equation.
[Expression 10]

  The weighting factor Ws in the equation [10] is a value indicating the ratio of each subband S contributing to the conspicuousness of the tile boundary, and may be determined based on a subjective evaluation experiment of the image. FIG. 7 is a table showing an example of the weighting factor Ws determined in this way. Further, A in the formula [10] is a proportionality constant.
[0059]
  In step S43, when the variance is calculated by dividing the tile into an upper side, a lower side, a right side, and a left side, it is preferable to obtain the evaluation parameter k for each of these categories.
[0060]
  Incidentally, although the evaluation parameter k is determined using the equation [10] here, the present invention is not limited to this evaluation equation. Any evaluation formula that calculates an evaluation parameter indicating that the tile boundary is not conspicuous as the variance value increases is sufficient.
[0061]
  FIG. 8 is a table showing the value of the evaluation parameter k by picking up a characteristic image area of the test image (person image). It can be predicted that the tile boundary is more conspicuous as the value of the evaluation parameter k is larger.
Step S45: Subsequently, the microprocessor 39 adjusts the value of the evaluation parameter k between adjacent tiles. Here, the microprocessor 39 first sets a local region including the upper, lower, left and right adjacent tiles for each tile to be processed. Next, the microprocessor 39 performs smoothing (simple average, weighted average, etc.) of the evaluation parameter k within the local region.
[0062]
  FIG. 9A is a map in which the evaluation parameters k of the person image are arranged in accordance with the tile sections. FIG. 9B is a map showing the evaluation parameter k after the smoothing process.
  Instead of the smoothing calculation here, calculation calculations such as a MAX value and a median value may be used.
[0063]
  In step S43, when the standard deviation is calculated by dividing the tile into an upper side, a lower side, a right side, and a left side, the evaluation parameter k may be adjusted between actually adjacent sides. (For example, the evaluation parameter k may be adjusted between the left side of the tile located on the right side and the right side of the tile located on the left side.)
[0064]
  By such adjustment of the evaluation parameter k, when there is a tile or a side where the tile boundary is conspicuous, it is determined that the tile boundary is also conspicuous for the tile or the tile in contact with the side.
Step S46: The microprocessor 39 considers such a prediction situation based on the evaluation parameter k, and determines the visual weight of each subband using the following equation.
γ3LL = 1.0
γ3LH = γ3HL = exp (-1 / 18 × k)
γ3HH ＝ exp (-2 / 18 × k)
γ2LH = γ2HL = exp (-4 / 18 × k)
γ2HH ＝ exp (-8 / 18 × k)
γ1LH ＝ γ1HL ＝ exp (-9 / 18 × k)
γ1HH ＝ exp (-18 / 18 × k)
  In the above equation, the visual weight γ3LL of the low frequency sub-band is set relatively large as the evaluation parameter k increases.
[0065]
  The microprocessor 39 transmits the visual weight thus selected to the quantization unit 23.
Step S47: The quantization unit 23 optimizes the quantization step of the transform coefficient for each tile so as to minimize the visual distortion of the quantization distortion according to the visual weight set for each tile. Since the procedure of this optimization method is the same as step S5 of the first embodiment, a duplicate description here is omitted.
Step S48: The quantization unit 23 quantizes the transform coefficient for each tile using the quantization step width of each subband optimized for each tile.
[0066]
  By the way, in the above-described optimization of the quantization step width, the quantization step width of the low frequency subband is set narrower as the evaluation parameter k is larger. As a result, as the tile has a larger evaluation parameter k, the compression distortion in the low band (here, quantization distortion) becomes smaller.
Steps S49 to S53: Same as steps S7 to S11 described in the first embodiment.
Step S54: The microprocessor 39 adds data relating to the quantization step width determined for each tile to the compressed data in preparation for data use at the time of image expansion, and generates a compressed file. The recording unit 32 records and saves the generated compressed file in the memory card 33.
[0067]
  The imaging operation and image compression processing of the electronic camera 11 are completed by the series of operations described above.
[0068]
[Effects of the third embodiment, etc.]
  By the operation described above, the same effect as that of the first embodiment can be obtained also in the third embodiment.
[0069]
  In particular, in the third embodiment, tiles whose tile boundaries are conspicuous (for example, tiles such as blue sky) are appropriately selected, and the low-frequency compression distortion (that is, generation of tile boundaries), which is a problem, can be suppressed in units of tiles. It becomes possible. Conversely, tiles whose tile boundaries are visually inconspicuous (for example, tiles with complex patterns) can be appropriately selected to allow moderately low-frequency compression distortion. As a result, for tiles with complex patterns, image information of mid- to high-frequency components can be compressed as accurately as possible to allow the generation of tile boundaries, and image information deterioration due to compression can be prevented as much as possible. become.
[0070]
  As described above, in the third embodiment, the frequency distribution of the compression code amount is adjusted for each tile on the basis of the conspicuousness of the tile boundary. As a result, at a very high compression rate where the tile boundary becomes a real problem, it is possible to realize adjustment that suppresses image information deterioration as much as possible while improving the tile boundary.
[0071]
  Furthermore, in the third embodiment, low-range compression distortion is suppressed also around the tile where the tile boundary is conspicuous. As a result, it is possible to improve the situation where low-frequency distortion occurs around the flat tile and the tile boundary becomes conspicuous as a result.
[0072]
  In the above-described embodiment, the conversion coefficient and the standard deviation of pixel values (square root of variance) are obtained to evaluate the conspicuousness of gradation shift. However, the present invention is not limited to this. In general, the statistical property of the present invention may be any amount that correlates with “easy to notice gradation shift”. For example, it may be a statistical property indicating gradation flatness, gradation uniformity, gradation monotone, and the like.
[0073]
  The “dispersion” of the present invention may be a scale indicating the degree of dispersion of values, and is not limited to the statistical calculation formula shown by the formula [9].
[0074]
  Incidentally, the dynamic range of the pixel value is a value that fluctuates greatly if there is a large gradation change even instantaneously. Therefore, the dynamic range of the pixel value is not a value indicating the flatness of the gradation, and is not an amount correlated with “easiness of gradation deviation”.
[0075]
  In the embodiment described above, the quantization step width of the transform coefficient is adjusted and set as the image compression parameter. However, the present invention is not limited to this. In general, the image compression parameter of the present invention may be any parameter that affects the appearance of compression distortion at the tile boundary. For example, the image compression parameter may be a parameter for adjusting the lower digit truncation of the bit plane, a parameter for adjusting the bit stream truncation, a parameter for adjusting the ROI encoding selection area, or the like.
[0076]
  In the first and second embodiments, the image compression parameters are uniformly changed for the entire plurality of tiles. In this case, since image compression is uniformly performed on the entire plurality of tiles, there is little risk of unevenness in the appearance of compression distortion on the screen. However, the present invention is not limited to this. For example, as shown in the third embodiment, the image compression parameter may be adjusted and set for each tile. Further, the image compression parameters may be individually adjusted and set at the tile boundary and inside the tile.
[0077]
  In the above-described embodiment, the case where the electronic camera 11 performs the image compression of the present invention has been described. However, the present invention is not limited to this. For example, the image compression apparatus of the present invention may be realized on a computer. In this case, an image compression program can be obtained by program-coding the image compression procedure in the above-described embodiment.TheCreate it.
[0078]
  The present invention can be implemented in various other forms without departing from the spirit or main features thereof. For this reason, the above-described embodiments are merely examples in all respects and should not be interpreted in a limited manner. The scope of the present invention is indicated by the scope of claims, and is not restricted by the text of the specification. Further, all modifications and changes belonging to the equivalent scope of the claims are within the scope of the present invention.
[0079]
[Possibility of industrial use]
  In the image compression according to the present invention, a statistical property correlating with “the visual conspicuousness of gradation shift” is obtained from the image data, and the compression is performed according to the conspicuousness of the tile boundary predicted by the statistical property. Adjust and set image compression parameters that affect the amount of distortion.
  Therefore, it is easy to take appropriate measures such as suppressing compression distortion in advance for image data in which tile boundaries are likely to be noticeable when the image is expanded.
[0080]
[Brief description of the drawings]
FIG. 1 is a diagram showing a configuration of an electronic camera 11 in the present embodiment.
FIG. 2 is a flowchart illustrating image compression processing according to the first embodiment.
FIG. 3 is a diagram illustrating a state of subband division by wavelet transform.
FIG. 4 is a flowchart illustrating image compression processing according to the second embodiment.
FIG. 5 is a flowchart illustrating image compression processing according to the third embodiment.
FIG. 6 is a diagram showing standard deviations (square roots of variance) of transform coefficients for each subband by picking up characteristic tiles of a test image.
FIG. 7 is a diagram illustrating an example of a weighting factor Ws for each subband used for calculating the evaluation parameter k.
FIG. 8 is a diagram showing an evaluation parameter k by picking up a characteristic tile of a test image.
FIG. 9 is a map in which evaluation parameters k are arranged according to tile divisions.

Claims (5)

An image compression apparatus that divides image data into a plurality of tiles and performs image compression for each tile,
A statistical processing unit that processes the image data and obtains a statistical property that correlates with “easiness of the gradation shift when the image data includes a gradation shift”;
Parameter setting that predicts the conspicuousness of tile boundaries during image expansion based on the statistical properties, and adjusts and sets “image compression parameters that affect the appearance of compression distortion” according to the conspicuousness of tile boundaries And
An image compression unit that compresses the image data using the adjusted image compression parameter, and
The statistical property is a variance obtained for a transform coefficient obtained by wavelet transforming the image data,
The image compression parameter is a quantization step width;
The parameter setting unit is configured to reduce the compression of the low-frequency spatial frequency component so that the variance of the weighting increases as the wavelet transform coefficient is moved around the tile for each sub-band , and the smaller the variance is, the lower the sub-band subbands are. An image compression apparatus characterized in that a quantization step width is set narrow . The image compression apparatus according to claim 1.
The statistical processing unit obtains the statistical property for each tile,
The parameter setting unit predicts the conspicuousness of the tile boundary in units of tiles based on the statistical property of the tile unit. An image compression apparatus , wherein the image compression parameter is adjusted and set in a reducing direction . The image compression apparatus according to claim 2 .
The image compression apparatus , wherein the parameter setting unit adjusts and sets the image compression parameter in a direction that reduces compression distortion of a low-frequency spatial frequency component even for a tile adjacent to a tile in which the tile boundary is conspicuous . An image compression program for causing a computer to function as the statistical processing unit, the parameter setting unit, and the image compression unit according to any one of claims 1 to 3. An image compression method that divides image data into a plurality of tiles and performs image compression for each tile,
A statistical processing step of processing the image data to obtain a statistical property that correlates with “easiness of the gradation shift when the image data has a gradation shift”;
Parameter setting that predicts the conspicuousness of tile boundaries during image expansion based on the statistical properties, and adjusts and sets “image compression parameters that affect the appearance of compression distortion” according to the conspicuousness of tile boundaries Steps,
An image compression step for compressing the image data using the image compression parameters that have been adjusted;
Have
The statistical property is a variance obtained for a transform coefficient obtained by wavelet transforming the image data,
The image compression parameter is a quantization step width;
The parameter setting unit is configured to reduce the compression of the low-frequency spatial frequency component so that the variance of the weighting increases as the wavelet transform coefficient is moved around the tile for each sub-band, and the smaller the variance is, the lower the sub-band subbands are. An image compression method characterized by setting a quantization step width to be narrow.

Images