JP3726333B2 - Quantization bit number conversion apparatus and method - Google Patents

Description

ｘ´：ｘ₄ の１０ビット予測値
ｘ_i ：８ビット入力画素値
ｗ_i ：予測係数
【００１３】
次に、重心法が用いられたクラス分類適応処理の構成例を図３に示す。この図３を用いて重心法を用いたクラス分類適応処理の概要を説明する。入力信号ｄ５は、入力端子６からクラス分類部７へ供給され、クラス分類部７では、入力信号ｄ５の波形特徴に基づくクラス分類が行われ、分類結果がｄ６として最適予測値ＲＯＭ８へ供給される。この最適予測値ＲＯＭ８では、クラス分類部７から供給された分類クラスｄ６に応答して最適予測値が読み出され、この最適予測値は、ｄ７として出力される。
【００１４】
ここで、この発明に係る量子化ビット数変換装置の第１の実施例を図４に示す。１１で示す入力端子を介して外部から画素データｄ１１が供給される。この画素データｄ１１は、放送などによる伝送あるいはビデオテープレコーダ装置などの再生によって、供給される映像信号が８ビットで２５６階調に量子化されている。この８ビット信号からなる画素データｄ１１は、処理単位のブロックに分割する画素データ分割手段であるブロック化回路１２へ供給される。図２に示すように（３×３）ブロックの９画素ｘ₀ 〜ｘ₈ にブロック化された画素データｄ１２およびｄ１４は、予測式演算回路１３およびＡＤＲＣ（Adaptive Dynamic Range Coding ）回路１４へ供給される。
【００１５】
ＡＤＲＣ回路１４は、ブロック毎の画素データｄ１４の２次元的なレベル分布のパターンを検出するために、このレベル分布のパターンを示すデータを圧縮してパターン圧縮データを出力する情報圧縮手段である。このＡＤＲＣ回路１４からのパターンｄ１５は、クラスコード発生回路１５へ供給される。クラスコード発生回路１５では、供給されたパターンｄ１５に基づいて、そのブロックの画素データが属するクラスが検出される。検出されたクラス情報であるクラスコードｄ１６は、予測係数メモリ１６へ供給される。予測係数メモリ１６では、外部から供給された８ビットの画素データを１０ビットで量子化された画素データへ変換するための情報である予測係数が予め求められて記憶されており、クラスコード発生回路１５からのクラスコードｄ１６に応答して予測係数ｄ１７が出力される。
【００１６】
この予測係数ｄ１７は、予測式演算回路１３へ供給される。予測式演算回路１３では、予測係数ｄ１７とブロック化された８ビットの画素データｄ１２とから式（１）に示す線形一次結合式の予測式に基づいた演算により１０ビットで量子化された画素データに変換される。この画素データｄ１３は、予測式演算回路１３からレベル制限回路１７へ供給される。レベル制限回路１７では、８ビットデータを１０ビットデータへ変換した際に周辺画素の影響により元々の存在区間を逸脱してしまうデータの補正が行われる。この補正の一例としてこのレベル制限回路１７では、存在区間を逸脱してしまうデータのクリッピングが行われる。
【００１７】
つまり、１０進法で表現して、例えばレベル１の８ビットデータの真値存在区間は、本来レベル４〜７の間に存在する１０ビットデータとなる。ところが予測演算により例えば３あるいは１０といったように真値存在区間外にデータが変換されてしまう場合もありえる。これは本来あるべき真値存在区間を逸脱しているのでレベル制限回路１７では、それぞれ４あるいは７へクリッピングが行われる。クリッピング処理が行われた画素データｄ１８は、出力端子１８から取り出される。
【００１８】
次に、この発明に係る量子化ビット数変換装置の第２の実施例を図５に示す。入力端子２１から供給される画素データｄ２１は、クラス分類適応処理部２２および後処理部２３へ供給される。クラス分類適応処理部２２は、クラス分類部２５、予測係数ＲＯＭ２６および予測演算部２７から構成される。後処理部２３は、判定部２８、クラス分類部２９、予測係数ＲＯＭ３０、予測演算部３１および選択部３２から構成される。クラス分類適応処理部２２へ供給された画素データｄ２１は、クラス分類部２５および予測演算部２７へ供給される。クラス分類部２５では、供給された画素データｄ２１から図２に示すように（３×３）ブロックの画素ｘ₀ 〜ｘ₈ が抽出され、注目画素をｘ₄ として、クラスが分類される。このクラスは、ｄ２２として予測係数ＲＯＭ２６へ供給される。
【００１９】
予測係数ＲＯＭ２６では、予め記憶された予測係数の中から供給されたクラスｄ２２に対応する９個の予測係数ｗ₀ 〜ｗ₈ が読み出される。図２に示した９個の画素ｘ₀ 〜ｘ₈ をそれぞれ１ビットで表現したクラスを用いた一例のため、５１２種類のクラスを有し、クラス毎に９個の予測係数ｗ₀ 〜ｗ₈ が記憶される。読み出された予測係数ｄ２３は、予測演算部２７へ供給される。予測演算部２７では、８ビットの画素データｄ２１と予測係数ｄ２３とを用いて、後述する線形１次結合式から１０ビットの予測値ｄ２４が生成される。生成された１０ビットの予測値ｄ２４は、クラス分類適応処理部２２の出力として後処理部２３へ供給される。
【００２０】
後処理部２３では、予測値ｄ２４が判定部２８、クラス分類部２９、予測係数ＲＯＭ３０、予測演算部３１および選択部３２へ供給され、入力端子２１からの画素データｄ２１が判定部２８へ供給される。クラス分類部２９では、図２に示す注目画素ｘ₄ を中心として１０ビットからなる９個の画素を用いてクラスが分類される。分類されたクラスは、ｄ２６として予測係数ＲＯＭ３０へ供給される。予測係数ＲＯＭ３０では、予め記憶された予測係数の中から供給されたクラスｄ２６に対応する９個の予測係数ｗ₀ 〜ｗ₈ が読み出される。この予測係数ＲＯＭ３０は、上述の予測係数ＲＯＭ２６と同様に５１２種類のクラスを有し、クラス毎に９個の予測係数ｗ₀ 〜ｗ₈ が記憶される。読み出された予測係数ｄ２７は、予測演算部３１へ供給される。予測演算部３１では、１０ビットの予測値ｄ２４と予測係数ｄ２７とを用いて、線形１次結合式から１０ビットの予測値ｄ２８が生成される。生成された１０ビットの予測値ｄ２８は、補正された１０ビットの予測値として、選択部３２へ供給される。
【００２１】
判定部２８では、入力端子２１からの画素データｄ２１と１０ビットの予測値ｄ２４とが供給され、画素データｄ２１の示す真値存在区間から予測値ｄ２４が外れているか否かが判定される。この判定結果ｄ２５は、選択部３２へ供給される。選択部３２では、判定結果ｄ２５に基づいてクラス分類適応処理部２２からの予測値ｄ２４を出力するか、後処理部２３からの補正された予測値ｄ２８を出力するかが選択される。具体的には、クラス分類適応処理部２２からの予測値ｄ２４が画素データｄ２１の示す真値存在区間から逸脱していないと判定されると、選択部３２では、予測値ｄ２４が選択され、予測値ｄ２４が画素データｄ２１の示す真値存在区間から逸脱していると判定されると、選択部３２では、補正された予測値ｄ２８が選択される。選択された予測値は、予測結果ｄ２９として出力端子２４から出力される。
【００２２】
ここで、この第２の実施例では、後処理部２３のクラス分類部２９および予測演算部３１に１０ビットの予測値ｄ２４が供給され、この予測値ｄ２４に基づいて予測値ｄ２８が生成されているが、入力端子２１からの画素データｄ２１をクラス分類部２９および予測演算部３１に供給して、この画素データｄ２１に基づいて予測値ｄ２８を生成しても良い。また、予測値ｄ２４と画像データｄ２１とをクラス分類部２９および予測演算部３１へ供給し、この２つのデータに基づいて予測値ｄ２８を生成しても良い。
【００２３】
次に、この発明に係る量子化ビット数変換装置の第３の実施例を図６に示す。入力端子４１から供給される画素データｄ３１は、クラス分類適応処理部４２および後処理部４３へ供給される。クラス分類適応処理部４２は、クラス分類部４５および最適予測値ＲＯＭ４６から構成される。後処理部４３は、判定部４７、クラス分類部４８、最適予測値ＲＯＭ４９および選択部５０から構成される。クラス分類適応処理部４２へ供給された画素データｄ２１は、クラス分類部４５へ供給される。クラス分類部４５では、供給された画素データｄ３１から図２に示すように注目画素ｘ₄ として、（３×３）ブロックの画素ｘ₀ 〜ｘ₈ が抽出され、クラスが分類される。このクラスは、ｄ３２として最適予測値ＲＯＭ４６へ供給される。
【００２４】
最適予測値ＲＯＭ４６では、予め記憶された最適予測値の中から供給されたクラスｄ３２に対応する１０ビットからなる最適予測値ｄ３３が読み出される。図２に示した９個の画素ｘ₀ 〜ｘ₈ をそれぞれ１ビットで表現したクラスを用いた一例のため、５１２種類のクラスを有し、クラス毎に最適予測値が記憶される。読み出された最適予測値ｄ３３は、後処理部４３へ供給される。後処理部４３では、最適予測値ｄ３３が判定部４７、クラス分類部４８、最適予測値ＲＯＭ４９および選択部５０へ供給され、入力端子４１からの画素データｄ３１が判定部４７へ供給される。
【００２５】
クラス分類部４８では、図２に示す注目画素ｘ₄ を中心として１０ビットからなる９個の画素を用いてクラスが分類される。分類されたクラスは、ｄ３５として最適予測値ＲＯＭ４９へ供給される。最適予測値ＲＯＭ４９では、予め記憶された最適予測値の中から供給されたクラスｄ３５に対応する１０ビットからなる最適予測値ｄ３６が読み出される。この最適予測値ＲＯＭ４９は、上述の最適予測値ＲＯＭ４６と同様に５１２種類のクラスを有し、クラス毎に最適予測値が記憶される。読み出された１０ビットの最適予測値ｄ３６は、補正された１０ビットの予測値として、選択部３２へ供給される。
【００２６】
判定部４７では、入力端子４１からの画素データｄ３１と１０ビットの最適予測値ｄ３３とが供給され、画素データｄ３１の示す真値存在区間から最適予測値ｄ３３が逸脱しているか否かが判定される。この判定結果ｄ３４は、選択部５０へ供給される。選択部５０では、判定結果ｄ３４に基づいて最適予測値ｄ３３を出力するか、最適予測値ｄ３３を出力するかが選択される。具体的には、クラス分類適応処理部４２からの最適予測値ｄ３３が画素データｄ３１の示す真値存在区間から逸脱していないと判定されると、選択部５０では、最適予測値ｄ３３が選択され、最適予測値ｄ３３が画素データｄ３１の示す真値存在区間から逸脱していると判定されると、選択部５０では、補正された最適予測値ｄ３６が選択される。選択された最適予測値は、予測結果ｄ３７として出力端子４４から出力される。
【００２７】
ここで、この第３の実施例では、後処理部４３のクラス分類部４８に１０ビットの最適予測値ｄ３３が供給され、この最適予測値ｄ３３に基づいて最適予測値ｄ３６が読み出されているが、入力端子４１からの画素データｄ３１をクラス分類部４８に供給して、この画素データｄ３１に基づいて最適予測値ｄ３６を読み出しても良い。また、最適予測値ｄ３３と画素データｄ３１とをクラス分類部４８へ供給し、この２つのデータに基づいて最適予測値ｄ３６を読み出しても良い。
【００２８】
さらに、上述の第２および第３の実施例を組み合わせ、例えばクラス分類適応処理部に予測法を用い、後処理部に重心法を用いる、また、クラス分類適応処理部に重心法を用い、後処理部に予測法を用いることも可能である。
【００２９】
このように、この発明は、クラス分類適応処理を用いた多ビットディジタル信号への信号変換法の一例として８ビットから１０ビットへの信号変換法を用いるものである。上述した図１１に示すように８ビット信号値と１０ビット信号値との対応図を考慮すると、クラス分類適応処理を用い１０ビット予測信号値は真値存在区間内に存在しなくてはならない。そこで、この区間を超える予測信号値に対し、再びクラス分類適応処理を適用することでより高性能な量子化ビット数変換を達成するものである。
【００３０】
ここで、上述した実施例におけるクラス分類適応処理部２２および４２で用いられているクラス分類適応処理を簡単に説明する。このクラス分類適応処理とは、入力信号の特徴に基づき入力信号をいくつかのクラスに分類し、予め用意された適切な適応処理をクラス毎に実行し所望の出力値を得る手法である。まず、クラス分類法の例としては、入力信号（８ビットＰＣＭデータ）に対して、クラス生成タップを設定し、入力信号のレベル分布のパターンによりクラスを生成する手法が挙げられる。信号波形のクラス生成法としては、次の例などが提案されている。
【００３１】
１）ＰＣＭ（Pluse Code Modulation ）データを直接使用する方法
２）ＡＤＲＣを適用する方法
３）ＤＰＣＭ（Differential PCM）を適用する方法
４）ＢＴＣ（Block Trancation Coding ）を適用する方法
５）ＶＱ（Vector Quantization ）を適用する方法
６）周波数領域クラス（ＤＣＴ（Discrete Cosine Transform ）、アダマール変換、フーリエ変換その他）を適用する方法
【００３２】
ＰＣＭデータを直接使用する場合、クラス分類用に８ビットデータを７タップ使用すると、２⁵⁶という膨大な数のクラスに分類される。信号波形の特徴を掴むという意味では理想的であるが、回路上の負担は大きく、実用上問題である。そこで実際は、ＡＤＲＣなどを適用しクラス数の削減を図る。ＡＤＲＣは、信号圧縮技術として開発された手法であるが、クラス表現に適している。基本的には、再量子化処理であり式（２）で示される。
【００３３】
ｑ_i ＝（ｘ_i −ＭＩＮ）／（ＤＲ／２^k ） （２）
ただし、ｑ_i ：ＡＤＲＣコード
ｘ_i ：入力画素値
ＭＩＮ：近傍領域内最小値
ＤＲ：近傍領域内ダイナミックレンジ
ｋ：再量子化ビット数
【００３４】
注目画素近傍の数タップに対し式（２）で定義されるＡＤＲＣを用いて生成されるＡＤＲＣコードに基づきクラス分類を行う。例えば、７画素データに対し、１ビットの再量子化を実行する１ビットＡＤＲＣを適用すると、７画素から定義されるダイナミックレンジに基づき、それらの最小値を除去した上で、７タップのデータを適応的に１ビット量子化する。その結果、７画素データを７ビットで表現することになり、１２８クラスに削減することが可能となる。他に上述した圧縮技術として一般的なものをクラス分類法として用いることが提案されている。また、クラス分類の性能を更に向上させるため、入力信号のアクティビティーも考慮した上でクラス分類が行われることがある。
【００３５】
アクティビティーの判定法としては、クラス分類法にＡＤＲＣを使用した場合、ダイナミックレンジを用いることが多い。また、ＤＰＣＭならば差分絶対値和、ＢＴＣのときは標準偏差の絶対値などが用いられる。このときには、アクティビティーによる分類結果毎にＡＤＲＣクラス分類などを行うことになる。また、学習過程において、アクティビティーの小さいデータを学習対象から外す。この理由は、アクティビティーの小さい部分は、ノイズの影響が大きく、本来のクラスの予測値から外れることが多い。それを学習に入れると予測精度が低下する。これを避けるため、学習においては、アクティビティーの小さいデータを除外する。
【００３６】
ここで、予測法で用いられる予測係数ＲＯＭからの予測係数は、予め学習により生成しておく。この学習方法について述べる。式（１）の線形一次結合モデルに基づく予測係数を最小自乗法により生成する一例を示す。最小自乗法は、次のように適用される。一般化した例として、Ｘを入力データ、Ｗを予測係数、Ｙを予測値として次の式（３）を考える。
【００３７】
観測方程式；ＸＷ＝Ｙ （３）
【数２】

【００３８】
上述の観測方程式により収集されたデータに最小自乗法を適用する。式（１）の例においては、ｎ＝９、ｍが学習データ数となる。式（３）の観測方程式をもとに、式（５）の残差方程式を考える。
【００３９】
残差方程式；
【数３】

【００４０】
式（５）の残差方程式から、各ｗ_i の最確値は、
【数４】

を最小にする条件が成り立つ場合と考えられる。すなわち、次の式（６）の条件を考慮すれば良いわけである。
【００４１】
【数５】

【００４２】
式（６）のｉに基づくｎ個の条件を考え、これを満たすｗ₁ 、ｗ₂ 、・・・ｗ_n を算出すれば良い。そこで、残差方程式（５）から式（７）が得られる。
【００４３】
【数６】

【００４４】
式（６）と式（３）により式（８）が得られる。
【００４５】
【数７】

【００４６】
そして、式（５）および式（８）から次の正規方程式（９）が得られる。
【００４７】
【数８】

【００４８】
式（９）の正規方程式は、未知数の数ｎと同じ数の方程式を立てることが可能であるので、各ｗ_i の最確値を求めることができる。そして、掃き出し法（Gauss-Jordanの消去法）を用いて連立方程式を解く。この連立方程式が解かれることよって、クラス毎に予測係数がＲＯＭなどの記憶媒体に格納される。この格納されたＲＯＭは、予測係数ＲＯＭとして使用される。
【００４９】
次に、上述した予測係数ＲＯＭに記憶される予測係数の学習の一例を図７のフローチャートを用いて説明する。このフローチャートは、ステップＳ１から学習処理の制御が始まり、ステップＳ１の学習データ形成では、例えば１フレームの中の８ビットの画素データとそれに対応する１０ビットの画素データとから学習データが形成される。フィールド内またはフレーム内の周辺画素の値が学習データとして採用される。注目画素の真値と複数の周辺画素の値とが一組の学習データである。
【００５０】
ここで、例えばＡＤＲＣを使用する場合、周辺画素で構成されるブロックのダイナミックレンジが所定のしきい値より小さいもの、すなわちアクティビティーの低いものは、学習データとして扱わない制御がなされる。アクティビティーが低いものは、ノイズの影響を受けやすく、正確な学習結果が得られないおそれがあるからである。ステップＳ２のデータ終了では、入力された全データ、例えば１フレームのデータの処理が終了していれば、ステップＳ５の予測係数決定へ制御が移り、終了していなければ、ステップＳ３のクラス決定へ制御が移る。
【００５１】
ステップＳ３のクラス決定は、上述のように、フィールド内またはフレーム内の所定の８ビットの画素データに基づいたクラス決定がなされる。ステップＳ４の正規方程式生成では、上述した式（９）の正規方程式が作成される。全データの処理が終了後、ステップＳ２のデータ終了から制御がステップＳ５に移る。このステップＳ５の予測係数決定では、この正規方程式を行列解法を用いて解いて、予測係数を決める。ステップＳ６の予測係数ストアでは、予測係数をメモリにストアし、この学習のフローチャートが終了する。
【００５２】
これらのように、小領域のブロック毎に分割された画像は、局所的な強い相関によってレベル的に近い値をとることが多い。このため、ＡＤＲＣを用いて、ブロック内の画素レベルの最大値および最小値を求めてブロック内ダイナミックレンジを定義することで、レベル方向の冗長度を大幅に除去することができる。例えば、図８に示すように８ビットの原データの持つ０〜２５５のダイナミックレンジの中で、各ブロック毎に再量子化するのに必要なブロック内ダイナミックレンジＡおよびＢは、大幅に小さくなることが分かる。このため、再量子化に必要なビット数は、大幅に低減することができる。
【００５３】
ここで、ＡＤＲＣを使用して、例えば８ビットデータを３ビットデータに圧縮処理すると、ブロック内ダイナミックレンジをＤＲ、ビット割当をｐ、ブロック内画素のデータレベルをｘ、再量子化コードをＱとして以下の式（１０）により図９Ａに示すようにブロック内の最大値ＭＡＸと最小値ＭＩＮとの間を指定されたビット長で均等に分割して再量子化を行う。
【００５４】
ＤＲ＝ＭＡＸ−ＭＩＮ＋１
Ｑ＝〔（ｘ−ＭＩＮ＋０．５）×２^p ／ＤＲ〕 （１０）
【数９】

【００５５】
ただし、
【数１０】

は、Ｑ＝ｉを満足するｘの平均を意味し、〔ｚ〕はｚ以下の最大の整数を表す。
【００５６】
次に、図９Ａに示すようにｐビット再量子化（この場合３ビット再量子化）で最上位の階調レベル（２^p −１）に相当するデータレベル内に存在するブロック内画素の平均値をとり、これを図９Ｂに示すように最大値ＭＡＸ´とする。また、図９Ａに示すように、この再量子化による最下位の階調レベル０に相当するデータレベル内に存在するブロック内画素の平均値をとり、これを図９Ｂにしめすように最小値ＭＩＮ´とする。
【００５７】
そして、新しく求められた最大値ＭＡＸ´および最小値ＭＩＮ´からブロック内ダイナミックレンジＤＲ´を新たに定義し直して、再量子化コードをｑとして、新しく求められたブロック内の最大値ＭＡＸ´および最小値ＭＩＮ´に基づいて式（１１）により図９Ｂに示すように再量子化を行う。
【００５８】
ＤＲ´＝ＭＡＸ´−ＭＩＮ´
ｑ＝〔（ｘ−ＭＩＮ´）×（２^p −１）／ＤＲ´＋０．５〕 （１１）
但し、〔ｚ〕は、ｚ以下の最大の整数を表す。
【００５９】
このように、ＡＤＲＣを用いて、二重の再量子化を行うことにより、画像の持つ局所的特徴としてブロック内ダイナミックレンジを定義し、主としてレベル方向の冗長度を適応的に除去することができ、ノイズの悪影響を受けることがなく、効率の良い情報量圧縮が行われる。
【００６０】
次に、ＡＤＲＣを用いて圧縮処理の行われた学習データは、学習データ毎にクラス分割される。すなわち、図２に示す注目画素ｘ₄ に注目した場合、８ビットデータｘ₀ 〜ｘ₈ をＡＤＲＣを適用してｐビットにデータ圧縮した結果の再量子化データをｑ₀ 〜ｑ₈ として、以下の式（１２）を用いてその学習データのクラス（ｃ）を算出する。
【００６１】
【数１１】

【００６２】
ここで、重心法による予測値の学習方法の一例となるフローチャートを図１０に示す。ステップＳ１１の初期化では、この学習を行うための準備として、クラスのデータテーブルＥ（＊）およびクラスの度数カウンタＮ（＊）へ０のデータが書き込まれる。ここで、“＊”は、全てのクラスを示し、データテーブルは、Ｅ（Ｃ０）となり、クラスＣ０に対応する度数カウンタは、Ｎ（Ｃ０）となる。ステップＳ１１の制御が終了すると、ステップＳ１２へ制御が移る。
【００６３】
ステップＳ１２のクラス検出では、学習対象画素の近傍データからクラスＣを決定する。例えば、上述の例のように注目画素を含む近傍８画素に１ビットＡＤＲＣを適用した場合、１２８クラスに分類される。また、このクラス分類の手法としては、上述のようにＡＤＲＣのほかにも、ＰＣＭ表現、ＤＰＣＭ、ＢＴＣ、ＶＱ、ＤＣＴ、アダマール変換などの分類法が考えられる。また、クラス分類対象データより構成されるブロックのアクティビティーを考慮する場合は、クラス数をアクティビティーによる分類の種類だけ増やしておくことも考えられる。
【００６４】
次に、ステップＳ１３のデータ検出では、目標とする教師信号ｅが検出される。ステップＳ１４のクラス別データ加算では、クラスＣ毎に教師信号ｅがそれぞれ加算され、ステップＳ１５のクラス別度数加算では、クラスＣの学習データの度数カウンタＮ（Ｃ）が＋１インクリメントされる。全学習対象データについて繰り返しステップＳ１２からステップＳ１５の制御が終了したか否かを判定するステップＳ１６では、全データの学習が終了していれば、ステップＳ１７へ制御が移り、全データの学習対象が終了していなければ、ステップＳ１２へ制御が移る。すなわち、ステップＳ１６は、全データの学習が終了になるまで、ステップＳ１２からステップＳ１５までの制御を繰り返し実行し、全てのクラスの度数カウンタＮ（＊）と対応する全てのクラスのデータテーブルＥ（＊）が生成される。
【００６５】
ステップＳ１７のクラス別平均値算出では、各クラスのデータテーブルＥ（＊）の内容であるデータ積算値を対応クラスの度数カウンタＮ（＊）の度数で、除算が実行され、各クラスの平均値が算出される。この処理は、教師信号分布の重心を算出することと等価である。この平均値が重心法による各クラスの最適予測値となる。そして、ステップＳ１８のクラス別平均値登録では、ＲＯＭなどの記憶手段に各クラスに対応する最適予測値を登録することで重心法による学習、すなわち、このフローチャートは、終了する。上述のように学習過程において、ノイズの影響を排除するため、アクティビティーの小さい場合を学習対象から外すことも考えられる。上述した予測法および重心法により、画素より詳細な位置となる画素値の予測を行うことができる。
【００６６】
次に、後処理部２３で用いられる予測係数ＲＯＭ３０の学習方法の一例を説明する。まず、上述したように１０ビットからなる教師信号から８ビットの信号が生成され、生成された８ビットの信号と１０ビットの教師信号から予測係数が得られる。この予測係数を用いて任意の８ビットの信号が１０ビットの信号へ変換される。そして、変換された１０ビットの信号と対応する１０ビットの教師信号との比較が行われる。その比較結果、１０ビットの信号は、１０ビットの教師信号と等しくなる値と、異なる値とへ判別される。このとき、異なる値と判別された１０ビットの教師信号と、変換された１０ビットの信号とを用いて、上述したように予測係数が求められる。このように求められた予測係数が、予測係数ＲＯＭ３０に使用される。
【００６７】
また、後処理部４３で用いられる最適予測値ＲＯＭ４９の学習方法も同様の手法が用いられる。具体的には、変換された１０ビットの予測値と１０ビットの教師信号との比較が行われ、１０ビットの教師信号と異なると判別された１０ビットの予測値は、異なる値と判別された１０ビットの教師信号と、変換された１０ビットの予測値とを用いて、上述したように最適な予測値が求められる。このように求められた予測値が最適予測値ＲＯＭ４９に使用される。
【００６８】
これらの実施例では、クラス分類に使用する画素と、予測演算に使用する画素とを同一のものとしても良いし、異なる画素を使用しても良い。
【００６９】
また、第２実施例では、クラス分類適応処理部と後処理部とで使用される予測係数ＲＯＭは、同一のものでも良いし、異なるもの、例えば予測係数の数が異なっても良い。
【００７０】
そして、第３の実施例では、クラス分類適応処理部と後処理部とで使用される最適予測値ＲＯＭは、同一のものでも良いし、異なるものでも良い。
【００７１】
【発明の効果】
この発明に依れば、クラス分類適応処理を用いているため、外部から供給された画素データを単に量子化ビット数を増加させるだけでなく、量子化ビット数の増加に見合う分の情報量が増加した信号に変換することができ、階調不足による擬似輪郭の発生などを防止することができる。
【００７２】
また、この発明に依れば、クラス分類適応処理を用いているため、入力されたディジタル信号より多いビット数のディジタル信号を生成することが容易にでき、生成されたディジタル信号の量子化雑音が低減される。このことによって、画像プロセスによる画質劣化が抑圧することができる。
【図面の簡単な説明】
【図１】この発明に適用される予測法の説明に用いるブロック図である。
【図２】この発明に適用される画素の位置を示した略線図である。
【図３】この発明に適用される重心法の説明い用いるブロック図である。
【図４】この発明が適用される量子化ビット数変換装置の第１の実施例である。
【図５】この発明が適用される量子化ビット数変換装置の第２の実施例である。
【図６】この発明が適用あれる量子化ビット数変換装置の第３の実施例である。
【図７】この発明に適用される予測法の学習方法の一例である。
【図８】この発明に適用されるＡＤＲＣの説明に用いる略線図である。
【図９】この発明に適用されるＡＤＲＣの説明に用いる略線図である。
【図１０】この発明に適用される重心法の学習の方法の一例である。
【図１１】量子化ビット数変換装置の説明に用いる略線図である。
【符号の説明】
１２・・・ブロック化回路、１３・・・予測式演算回路、１４・・・ＡＤＲＣ回路、１５・・・クラスコード発生回路、１６・・・予測係数メモリ、１７・・・レベル制御回路[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a quantization bit number conversion device suitable for use in, for example, a digital video tape recorder device (digital VTR), and in particular, converts an externally supplied image signal into an image signal having a larger quantization bit number. The present invention relates to an apparatus and a method for converting the number of quantized bits of an image signal to be output.
[0002]
[Prior art]
Today, the trend of digitization in the video field has been steadily steadily spreading, and some image exchanges using digital signals have already been completed and put into practical use. One of them is CCIR Rec. 601 and the like. This defines the format of Y / U / V digital component signals, and each pixel is defined by an 8-bit digital signal. Later, there is a growing demand for higher image quality, image quality degradation due to insufficient gradation due to 8-bit quantization during special effect processing such as image composition, deformation, and enlargement / reduction, or various signal processing in various video devices Due to the demands of the image process, such as ensuring the calculation accuracy of 10 bits, it was necessary to define each pixel with 10 bits, and the signal standard for 10-bit data was also determined. One example is SMPTE 259M, which is a serial digital interface. Therefore, when exchanging signals between different signal standards, Rec. It is necessary to convert an 8-bit signal defined by 601 or the like into a 10-bit signal.
[0003]
Thus, one of the necessary techniques for signal transfer between different digital signal formats is quantization bit number conversion. Here, as an example, consider the conversion from an 8-bit digital signal to a 10-bit digital signal. Figure shows the relationship of the signal values 11 Shown in In this example, the 8-bit digital signal value Q8 is a quantized representative value, and the original analog signal value (true value) of the 8-bit signal value Q8 is included in the true value existence section in the figure. Therefore, in order to convert an 8-bit digital signal into a 10-bit digital signal, the 8-bit signal value Q8 is converted into a 10-bit digital signal value Q10. ₀ ~ Q10 _Three One of the four types is selected and output. An example of conversion from a general 8-bit signal value to a 10-bit signal value is to add a certain code such as `00 'or` 01' to the lower 2 bits. As a result, the 10-bit signal value Q10 ₀ Will always be output.
[0004]
[Problems to be solved by the invention]
However, since the above-described quantized bit number conversion device simply increases the data length of each data from 8 bits to 10 bits, the output signal is the data encoded with 8-bit encoded data from the beginning. The 10-bit signal value Q10 is not changed. ₀ Reflects the attribute of the 8-bit signal value, so in the processing of DVE (Digital Video Effector), chroma key, switcher, etc. in the image process, if a certain signal level width is expanded, image quality degradation is noticeable due to quantization noise. Thus, there is a problem that a pseudo contour is generated due to insufficient gradation.
[0005]
Therefore, an object of the present invention is made in view of the above-mentioned problems, and it is not only to increase the number of quantization bits, but also to classify a signal with a small number of quantization bits using a class classification adaptive process. An object of the present invention is to provide an apparatus and method for converting the number of quantization bits that can be converted into a signal corresponding to an increase in quantity.
[0006]
[Means for Solving the Problems]
According to the first aspect of the present invention, a level distribution pattern of an input digital signal is detected for each block composed of a plurality of pixels of an input digital signal having a plurality of pixels, and an input is performed based on the detected pattern. Class classification means for determining the class to which a digital signal block belongs and outputting the class information, and storing conversion information for converting to a digital signal quantized with a larger number of quantization bits than the input digital signal for each class Storage means for outputting conversion information according to class information from the class classification means, conversion means for converting an input digital signal into a digital signal based on the output conversion information, and a quantum of the input digital signal. Detects whether the quantized value of the digital signal is included in the existing interval indicated by the quantized value, and the quantized value of the digital signal is Quantization bit number conversion, characterized by having a limiting means for clipping the quantized value of the digital signal so that the quantized value of the digital signal falls within the existing period when it is determined that the current value deviates from the present period Device.
In the invention according to claim 2, a plurality of pixels are arranged. Input Digital signal Consists of multiple pixels A level distribution pattern of the first digital signal is detected for each block, and based on the detected pattern, the first class to which the first digital signal block belongs is determined and the first class information is output. First class classification means for storing the first conversion information for converting into a second digital signal quantized with a larger number of quantization bits than the first digital signal is stored for each first class. First storage means for outputting the first conversion information in accordance with the first class information from the first class classification means, and a first digital signal based on the output first conversion information. For each block of the first conversion means for converting the signal into the second digital signal and the converted second digital signal, a level distribution pattern of the second digital signal is detected, and the detected pattern is based on the detected pattern. Second class classification means for determining a second class to which a block of the second digital signal belongs and outputting second class information; and a third digital signal closer to a true value for the second digital signal. Second conversion information for conversion into a signal is stored for each second class, and second conversion information is output in accordance with the second class information from the second class classification means. Storage means, second conversion means for converting the second digital signal into a third digital signal based on the output second conversion information, and an existing section indicated by the quantized value of the first digital signal Determining means for determining whether or not the quantized value of the second digital signal converted by the first converting means is included, and a second obtained from the first converting means based on the determination result The quantized value of the digital signal is If the second digital signal obtained from the first conversion means is selected and the quantized value of the second digital signal obtained from the first conversion means deviates from the existing section, the second conversion is performed. A quantizing bit number conversion device comprising: selecting means for selecting a third digital signal obtained from the means.
Furthermore, the invention according to claim 4 includes a plurality of pixels. Input Digital signal Consists of multiple pixels A level distribution pattern of the first digital signal is detected for each block, and based on the detected pattern, the first class to which the first digital signal block belongs is determined and the first class information is output. First class classifying means and a second digital signal quantized with a larger number of quantization bits than the first digital signal are stored for each first class, and from the first class classifying means, A first storage means for outputting a second digital signal in accordance with the first class information, and detecting a level distribution pattern of the second digital signal for each block of the output second digital signal. A second class classification means for determining a second class to which a block of the second digital signal belongs based on the detected pattern and outputting second class information; and a second digit A third digital signal closer to the true value than the second signal is stored for each second class, and a second digital signal is output according to the second class information from the second class classification means. And a determination unit that determines whether or not the existence section indicated by the quantization value of the first digital signal includes the quantization value of the second digital signal obtained from the first storage unit; If the quantized value of the second digital signal obtained from the first storage means is included in the existence section based on the determination result, the second digital signal obtained from the first storage means is selected, and the first And a selecting means for selecting a third digital signal obtained from the second storage means when the quantized value of the second digital signal obtained from the storage means deviates from the existing section. Bit number conversion It is the location.
[0007]
The invention according to claim 5 detects the level distribution pattern of the input digital signal for each block composed of a plurality of pixels of the input digital signal having a plurality of pixels, and inputs based on the detected pattern. A class classification step that determines the class to which the block of the digital signal belongs and outputs the class information, and stores conversion information for converting to a digital signal quantized with a larger number of quantization bits than the input digital signal for each class A step of outputting conversion information from the storage means according to the class information from the class classification step, a conversion step of converting the input digital signal to a digital signal based on the output conversion information, Whether the quantized value of the digital signal is included in the existing interval indicated by the quantized value of the input digital signal And a limiting step for clipping the quantized value of the digital signal so that the quantized value of the digital signal falls within the existing section when it is determined that the quantized value of the digital signal is out of the existing section. This is a quantization bit number conversion method characterized by the following.
In the invention according to claim 6, a plurality of pixels are arranged. Input Digital signal Consists of multiple pixels A level distribution pattern of the first digital signal is detected for each block, and based on the detected pattern, the first class to which the first digital signal block belongs is determined and the first class information is output. And first conversion information for converting to a second digital signal quantized with a larger number of quantization bits than the first digital signal is provided for each first class. And outputting the first conversion information from the first storage means according to the first class information from the first class classification step, and the output first conversion information A first conversion step for converting the first digital signal into a second digital signal, and a level distribution of the second digital signal for each block of the converted second digital signal. A second class classification step of detecting a turn, determining a second class to which a block of the second digital signal belongs based on the detected pattern and outputting second class information; and a second digital Second conversion information for converting the signal into a third digital signal closer to the true value is stored in the second storage means for each second class, and the second conversion information from the second class classification step is stored. Outputting the second conversion information from the second storage means according to the class information, and converting the second digital signal to the third digital signal based on the output second conversion information. And a determination step for determining whether or not the existence section indicated by the quantization value of the first digital signal includes the quantization value of the second digital signal converted in the first conversion step. When, If the quantized value of the second digital signal obtained from the first conversion step is included in the existing section based on the result of the determination, the second digital signal obtained from the first conversion step is selected, And a selection step for selecting a third digital signal obtained from the second conversion step when the quantized value of the second digital signal obtained from the conversion step deviates from the existing section. Bit number conversion method.
Furthermore, in the invention according to claim 7, a plurality of pixels are arranged. Input Digital signal Consists of multiple pixels A level distribution pattern of the first digital signal is detected for each block, and based on the detected pattern, the first class to which the first digital signal block belongs is determined and the first class information is output. A first class classification step, and a second digital signal quantized with a larger number of quantization bits than the first digital signal is stored in the first storage means for each first class, Outputting a second digital signal from the first storage means according to the first class information from the one class classification step, and a second digital signal for each block of the output second digital signal. The level distribution pattern is detected, and based on the detected pattern, the second class to which the second digital signal block belongs is determined and the second class information is output. A second digital classifying step and a third digital signal closer to the true value than the second digital signal are stored in the second storage means for each second class, and the second digital class signal from the second class classifying step Outputting the third digital signal from the second storage means in accordance with the class information of the second digital signal obtained from the first storage means in the existing section indicated by the quantized value of the first digital signal A determination step for determining whether or not a quantized value of the signal is included; and a quantized value of the second digital signal obtained from the first storage means based on the determination result is included in the existing section, When the second digital signal obtained from the first storage means is selected and the quantized value of the second digital signal obtained from the first storage means deviates from the existing section, the second digital signal obtained from the second storage means is obtained. Three Is the number of quantization bits conversion method characterized in that it comprises a selection step of selecting a Ijitaru signal.
[0008]
In the quantization bit number conversion apparatus and method according to the present invention, classification is performed based on a level distribution pattern of input 8-bit pixel data, and prediction coefficients corresponding to the class are read out and input. The 8-bit pixel data and the prediction coefficient are converted into a 10-bit prediction value, and classification is performed based on the converted 10-bit prediction value and the level distribution pattern of the input pixel data. The prediction coefficient is read, and 10-bit pixel data is generated from the input 8-bit pixel data and the prediction coefficient. Then, it is detected whether or not the generated 10-bit pixel data is included in the true value existence section of the input 8-bit pixel data. If not included, the predicted value is corrected. The As a correction, a 10-bit pixel data predicted from the 8-bit pixel data is output using a method for limiting the upper limit or the lower limit, or a method for performing classification adaptation processing using the prediction method and the centroid method. .
[0009]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, an embodiment of a quantization bit number conversion apparatus according to the present invention will be described in detail with reference to the drawings. First, in this embodiment, adaptive processing is performed for each classified class. A prediction method using a product-sum operation with a prediction coefficient prepared in advance and a centroid method for outputting a prediction value prepared in advance are provided. Proposed.
[0010]
FIG. 1 shows a configuration example of the class classification adaptive processing in which the prediction method is first used. The outline of the classification adaptation process using the prediction method will be described with reference to FIG. The input signal d0 is supplied from the input terminal to the class classification unit 1 and the prediction unit 2. The class classification unit 1 performs class classification based on the waveform characteristics of the input signal d0, and supplies the classification result to the prediction unit 2 as d1. The prediction unit 2 includes a prediction coefficient ROM 3 and a prediction calculation unit 4. In the prediction coefficient ROM 3, a prediction coefficient is read in response to the classification class d1 supplied from the class classification unit 1, and this prediction coefficient is supplied to the prediction calculation unit 4 as d2. In the prediction calculation unit 4, a prediction tap is formed from the input signal d0, and the product-sum calculation of Expression (1) is executed using the prediction coefficient given by the prediction coefficient d2. In this way, the 10-bit signal prediction value for the input 8-bit signal is output as d3.
[0011]
In this prediction method, as shown in the example of FIG. 2, 3 pixels × 3 lines (hereinafter referred to as (3 × 3) blocks), that is, a target pixel x _Four Neighboring 8-bit input pixels x ₀ ~ X ₈ A prediction tap is formed from 9 pixels, and an 8-bit pixel x _Four 10-bit signal values are predicted. The predicted value is shown in Formula (1).
[0012]
[Expression 1]

x ': x _Four 10-bit predicted value
x _i : 8-bit input pixel value
w _i : Prediction coefficient
[0013]
Next, FIG. 3 shows a configuration example of the class classification adaptive processing using the centroid method. The outline of the classification adaptation process using the center of gravity method will be described with reference to FIG. The input signal d5 is supplied from the input terminal 6 to the class classification unit 7. The class classification unit 7 performs class classification based on the waveform characteristics of the input signal d5, and supplies the classification result to the optimum predicted value ROM 8 as d6. . In the optimum prediction value ROM 8, the optimum prediction value is read in response to the classification class d6 supplied from the class classification unit 7, and this optimum prediction value is output as d7.
[0014]
FIG. 4 shows a first embodiment of the quantization bit number conversion apparatus according to the present invention. Pixel data d11 is supplied from the outside via an input terminal indicated by reference numeral 11. In the pixel data d11, the supplied video signal is quantized to 256 gradations with 8 bits by transmission by broadcasting or reproduction by a video tape recorder device or the like. The pixel data d11 composed of the 8-bit signal is supplied to the block forming circuit 12 which is a pixel data dividing means for dividing the block into processing units. As shown in FIG. 2, 9 pixels x (3 × 3) blocks ₀ ~ X ₈ The pixel data d12 and d14 that have been blocked into blocks are supplied to a predictive expression arithmetic circuit 13 and an ADRC (Adaptive Dynamic Range Coding) circuit 14.
[0015]
The ADRC circuit 14 is information compression means for compressing data indicating the level distribution pattern and outputting pattern compression data in order to detect a two-dimensional level distribution pattern of the pixel data d14 for each block. The pattern d15 from the ADRC circuit 14 is supplied to the class code generation circuit 15. The class code generation circuit 15 detects a class to which the pixel data of the block belongs based on the supplied pattern d15. The detected class code d16 is supplied to the prediction coefficient memory 16. In the prediction coefficient memory 16, a prediction coefficient which is information for converting 8-bit pixel data supplied from the outside into pixel data quantized with 10 bits is obtained in advance and stored. In response to the class code d16 from 15, the prediction coefficient d17 is output.
[0016]
The prediction coefficient d17 is supplied to the prediction formula calculation circuit 13. In the prediction formula calculation circuit 13, the pixel data quantized with 10 bits from the prediction coefficient d17 and the blocked 8-bit pixel data d12 by calculation based on the prediction formula of the linear linear combination formula shown in Formula (1). Is converted to The pixel data d13 is supplied from the prediction formula calculation circuit 13 to the level limiting circuit 17. The level limiting circuit 17 corrects data that deviates from the original existing section due to the influence of surrounding pixels when converting 8-bit data to 10-bit data. As an example of this correction, the level limiting circuit 17 performs clipping of data that deviates from the existing section.
[0017]
That is, expressed in decimal notation, for example, the true value existence section of level 1 8-bit data becomes 10-bit data originally existing between levels 4-7. However, data may be converted outside the true value existence section, for example, 3 or 10 by the prediction calculation. Since this deviates from the true value existence section which should be originally, the level limiting circuit 17 performs clipping to 4 or 7, respectively. The pixel data d18 that has been subjected to the clipping process is taken out from the output terminal 18.
[0018]
Next, FIG. 5 shows a second embodiment of the quantization bit number conversion apparatus according to the present invention. Pixel data d21 supplied from the input terminal 21 is supplied to the class classification adaptive processing unit 22 and the post-processing unit 23. The class classification adaptive processing unit 22 includes a class classification unit 25, a prediction coefficient ROM 26, and a prediction calculation unit 27. The post-processing unit 23 includes a determination unit 28, a class classification unit 29, a prediction coefficient ROM 30, a prediction calculation unit 31, and a selection unit 32. The pixel data d21 supplied to the class classification adaptive processing unit 22 is supplied to the class classification unit 25 and the prediction calculation unit 27. In the class classification unit 25, as shown in FIG. 2, the pixel x of the (3 × 3) block from the supplied pixel data d21. ₀ ~ X ₈ Is extracted, and the pixel of interest is x _Four The class is classified as This class is supplied to the prediction coefficient ROM 26 as d22.
[0019]
In the prediction coefficient ROM 26, nine prediction coefficients w corresponding to the class d22 supplied from the prediction coefficients stored in advance. ₀ ~ W ₈ Is read out. Nine pixels x shown in FIG. ₀ ~ X ₈ Is an example using a class in which each is represented by 1 bit, and has 512 kinds of classes, and nine prediction coefficients w for each class. ₀ ~ W ₈ Is memorized. The read prediction coefficient d23 is supplied to the prediction calculation unit 27. The prediction calculation unit 27 uses the 8-bit pixel data d21 and the prediction coefficient d23 to generate a 10-bit predicted value d24 from a linear linear combination formula described later. The generated 10-bit predicted value d24 is supplied to the post-processing unit 23 as an output of the class classification adaptive processing unit 22.
[0020]
In the post-processing unit 23, the predicted value d24 is supplied to the determination unit 28, the class classification unit 29, the prediction coefficient ROM 30, the prediction calculation unit 31, and the selection unit 32, and the pixel data d21 from the input terminal 21 is supplied to the determination unit 28. The In the class classification unit 29, the target pixel x shown in FIG. _Four The class is classified using 9 pixels of 10 bits centering on. The classified class is supplied to the prediction coefficient ROM 30 as d26. In the prediction coefficient ROM 30, nine prediction coefficients w corresponding to the class d26 supplied from the prediction coefficients stored in advance are stored. ₀ ~ W ₈ Is read out. The prediction coefficient ROM 30 has 512 types of classes, similar to the prediction coefficient ROM 26 described above, and nine prediction coefficients w for each class. ₀ ~ W ₈ Is memorized. The read prediction coefficient d27 is supplied to the prediction calculation unit 31. The prediction calculation unit 31 uses the 10-bit prediction value d24 and the prediction coefficient d27 to generate a 10-bit prediction value d28 from the linear linear combination formula. The generated 10-bit predicted value d28 is supplied to the selection unit 32 as a corrected 10-bit predicted value.
[0021]
The determination unit 28 is supplied with the pixel data d21 and the 10-bit predicted value d24 from the input terminal 21, and determines whether or not the predicted value d24 is out of the true value existing section indicated by the pixel data d21. The determination result d25 is supplied to the selection unit 32. The selection unit 32 selects whether to output the predicted value d24 from the class classification adaptive processing unit 22 or to output the corrected predicted value d28 from the post-processing unit 23 based on the determination result d25. Specifically, when it is determined that the predicted value d24 from the class classification adaptive processing unit 22 has not deviated from the true value existence section indicated by the pixel data d21, the selection unit 32 selects the predicted value d24 and performs prediction. If it is determined that the value d24 deviates from the true value existing section indicated by the pixel data d21, the selection unit 32 selects the corrected predicted value d28. The selected prediction value is output from the output terminal 24 as the prediction result d29.
[0022]
Here, in the second embodiment, a 10-bit prediction value d24 is supplied to the class classification unit 29 and the prediction calculation unit 31 of the post-processing unit 23, and a prediction value d28 is generated based on the prediction value d24. However, the pixel data d21 from the input terminal 21 may be supplied to the class classification unit 29 and the prediction calculation unit 31, and the prediction value d28 may be generated based on the pixel data d21. Further, the predicted value d24 and the image data d21 may be supplied to the class classification unit 29 and the prediction calculation unit 31, and the predicted value d28 may be generated based on these two data.
[0023]
Next, FIG. 6 shows a third embodiment of the quantization bit number conversion apparatus according to the present invention. Pixel data d31 supplied from the input terminal 41 is supplied to the class classification adaptive processing unit 42 and the post-processing unit 43. The class classification adaptation processing unit 42 includes a class classification unit 45 and an optimum predicted value ROM 46. The post-processing unit 43 includes a determination unit 47, a class classification unit 48, an optimum predicted value ROM 49, and a selection unit 50. The pixel data d21 supplied to the class classification adaptive processing unit 42 is supplied to the class classification unit 45. In the class classification unit 45, as shown in FIG. _Four (3 × 3) block of pixels x ₀ ~ X ₈ Are extracted and the classes are classified. This class is supplied to the optimum predicted value ROM 46 as d32.
[0024]
In the optimum prediction value ROM 46, the optimum prediction value d33 consisting of 10 bits corresponding to the class d32 supplied from among the optimum prediction values stored in advance is read out. Nine pixels x shown in FIG. ₀ ~ X ₈ Is an example using classes each expressed by 1 bit, and has 512 types of classes, and the optimum prediction value is stored for each class. The read optimum prediction value d33 is supplied to the post-processing unit 43. In the post-processing unit 43, the optimum predicted value d33 is supplied to the determining unit 47, the class classification unit 48, the optimum predicted value ROM 49, and the selecting unit 50, and the pixel data d31 from the input terminal 41 is supplied to the determining unit 47.
[0025]
In the class classification unit 48, the target pixel x shown in FIG. _Four The class is classified using 9 pixels of 10 bits centering on. The classified class is supplied to the optimum predicted value ROM 49 as d35. In the optimum prediction value ROM 49, the optimum prediction value d36 consisting of 10 bits corresponding to the class d35 supplied from among the optimum prediction values stored in advance is read out. This optimal prediction value ROM 49 has 512 types of classes, similar to the above-described optimal prediction value ROM 46, and the optimal prediction value is stored for each class. The read 10-bit optimum prediction value d36 is supplied to the selection unit 32 as a corrected 10-bit prediction value.
[0026]
The determination unit 47 is supplied with the pixel data d31 from the input terminal 41 and the 10-bit optimum prediction value d33, and determines whether or not the optimum prediction value d33 deviates from the true value existing section indicated by the pixel data d31. The The determination result d34 is supplied to the selection unit 50. The selection unit 50 selects whether to output the optimal prediction value d33 or the optimal prediction value d33 based on the determination result d34. Specifically, when it is determined that the optimum predicted value d33 from the class classification adaptive processing unit 42 has not deviated from the true value existence section indicated by the pixel data d31, the selecting unit 50 selects the optimum predicted value d33. When it is determined that the optimum predicted value d33 deviates from the true value existing section indicated by the pixel data d31, the selection unit 50 selects the corrected optimum predicted value d36. The selected optimal prediction value is output from the output terminal 44 as the prediction result d37.
[0027]
Here, in the third embodiment, the 10-bit optimum prediction value d33 is supplied to the class classification unit 48 of the post-processing unit 43, and the optimum prediction value d36 is read based on the optimum prediction value d33. However, the pixel data d31 from the input terminal 41 may be supplied to the class classification unit 48, and the optimum predicted value d36 may be read based on the pixel data d31. Alternatively, the optimum predicted value d33 and the pixel data d31 may be supplied to the class classification unit 48, and the optimum predicted value d36 may be read based on these two data.
[0028]
Further, the second and third embodiments described above are combined. For example, the prediction method is used for the class classification adaptive processing unit, the centroid method is used for the post-processing unit, and the centroid method is used for the class classification adaptive processing unit. It is also possible to use a prediction method for the processing unit.
[0029]
As described above, the present invention uses a signal conversion method from 8 bits to 10 bits as an example of a signal conversion method into a multi-bit digital signal using the class classification adaptive processing. Figure above 11 Considering the correspondence diagram between the 8-bit signal value and the 10-bit signal value as shown in FIG. 9, the 10-bit predicted signal value must exist in the true value existence section using the class classification adaptive processing. Therefore, higher-performance quantization bit number conversion is achieved by applying the class classification adaptive process again to the predicted signal value exceeding this interval.
[0030]
Here, the class classification adaptation processing used in the class classification adaptation processing units 22 and 42 in the embodiment described above will be briefly described. This class classification adaptive processing is a method of classifying an input signal into several classes based on the characteristics of the input signal, and executing appropriate adaptive processing prepared in advance for each class to obtain a desired output value. First, as an example of the class classification method, there is a method in which a class generation tap is set for an input signal (8-bit PCM data) and a class is generated based on a level distribution pattern of the input signal. The following examples have been proposed as signal waveform class generation methods.
[0031]
1) Direct use of PCM (Pluse Code Modulation) data
2) Method of applying ADRC
3) Method of applying DPCM (Differential PCM)
4) Method of applying BTC (Block Trancation Coding)
5) Method of applying VQ (Vector Quantization)
6) Method of applying frequency domain classes (DCT (Discrete Cosine Transform), Hadamard transform, Fourier transform, etc.)
[0032]
When PCM data is used directly, if 7 taps of 8-bit data are used for classification, 2 ⁵⁶ It is classified into a huge number of classes. Although it is ideal in terms of grasping the characteristics of the signal waveform, the burden on the circuit is large, which is a problem in practical use. Therefore, in actuality, ADRC or the like is applied to reduce the number of classes. ADRC is a technique developed as a signal compression technique, but is suitable for class expression. Basically, it is a re-quantization process and is represented by equation (2).
[0033]
q _i = (X _i -MIN) / (DR / 2 ^k (2)
However, q _i : ADRC code
x _i : Input pixel value
MIN: Minimum value in the neighborhood
DR: Dynamic range in the vicinity
k: Number of requantization bits
[0034]
Class classification is performed based on an ADRC code generated using ADRC defined by Expression (2) for several taps near the pixel of interest. For example, when 1-bit ADRC that performs 1-bit requantization is applied to 7-pixel data, the minimum value is removed based on the dynamic range defined from 7 pixels, and 7-tap data is converted. 1-bit quantization is adaptively performed. As a result, 7 pixel data is expressed by 7 bits, and can be reduced to 128 classes. In addition, it has been proposed to use a general compression technique described above as a class classification method. In order to further improve the performance of class classification, class classification may be performed in consideration of input signal activity.
[0035]
As an activity determination method, when ADRC is used as a classification method, a dynamic range is often used. Further, the sum of absolute differences is used for DPCM, and the absolute value of standard deviation is used for BTC. At this time, ADRC class classification is performed for each classification result by activity. Also, in the learning process, data with small activity is excluded from the learning target. The reason for this is that the small part of the activity is greatly affected by noise and often deviates from the predicted value of the original class. When it is put into learning, the prediction accuracy decreases. To avoid this, data with low activity is excluded in learning.
[0036]
Here, the prediction coefficient from the prediction coefficient ROM used in the prediction method is generated in advance by learning. This learning method is described. An example of generating a prediction coefficient based on the linear linear combination model of Expression (1) by the method of least squares is shown. The least squares method is applied as follows. As a generalized example, the following formula (3) is considered, where X is input data, W is a prediction coefficient, and Y is a predicted value.
[0037]
Observation equation; XW = Y (3)
[Expression 2]

[0038]
Apply the least squares method to the data collected by the above observation equation. In the example of Expression (1), n = 9 and m is the number of learning data. Consider the residual equation of equation (5) based on the observation equation of equation (3).
[0039]
Residual equation;
[Equation 3]

[0040]
From the residual equation of equation (5), each w _i The most probable value of is
[Expression 4]

It is considered that the condition for minimizing is satisfied. That is, the condition of the following formula (6) should be considered.
[0041]
[Equation 5]

[0042]
Consider n conditions based on i in Equation (6), and satisfy w ₁ , W ₂ ... w _n May be calculated. Therefore, the equation (7) is obtained from the residual equation (5).
[0043]
[Formula 6]

[0044]
Equation (8) is obtained from Equation (6) and Equation (3).
[0045]
[Expression 7]

[0046]
Then, the following normal equation (9) is obtained from the equations (5) and (8).
[0047]
[Equation 8]

[0048]
Since the normal equation of equation (9) can establish the same number of equations as the unknown number n, each w _i The most probable value of can be obtained. Then, the simultaneous equations are solved by using the sweep-out method (Gauss-Jordan elimination method). By solving the simultaneous equations, the prediction coefficient for each class is stored in a storage medium such as a ROM. This stored ROM is used as a prediction coefficient ROM.
[0049]
Next, an example of learning of the prediction coefficient stored in the above-described prediction coefficient ROM will be described with reference to the flowchart of FIG. In this flowchart, control of learning processing starts from step S1, and in the learning data formation in step S1, learning data is formed from, for example, 8-bit pixel data in one frame and corresponding 10-bit pixel data. . The values of peripheral pixels in the field or frame are adopted as learning data. The true value of the pixel of interest and the values of a plurality of surrounding pixels are a set of learning data.
[0050]
Here, for example, when ADRC is used, control is performed in which the dynamic range of a block composed of neighboring pixels is smaller than a predetermined threshold, that is, a low activity is not treated as learning data. This is because those with low activity are susceptible to noise and may not provide accurate learning results. At the end of the data in step S2, if the processing of all input data, for example, one frame of data has been completed, the control shifts to the prediction coefficient determination in step S5, and if not completed, the control proceeds to the class determination in step S3. Control is transferred.
[0051]
The class determination in step S3 is performed based on predetermined 8-bit pixel data in the field or frame as described above. In the generation of the normal equation in step S4, the normal equation of the above equation (9) is created. After the processing of all data is completed, control is transferred to step S5 from the end of data in step S2. In the prediction coefficient determination in step S5, this normal equation is solved using a matrix solution method to determine the prediction coefficient. In the prediction coefficient store in step S6, the prediction coefficient is stored in the memory, and the learning flowchart ends.
[0052]
As described above, an image divided for each block in a small region often takes a value close to the level due to a strong local correlation. For this reason, by using ADRC to determine the maximum value and the minimum value of the pixel level in the block and defining the dynamic range in the block, the redundancy in the level direction can be largely removed. For example, as shown in FIG. 8, in the dynamic range of 0 to 255 of the 8-bit original data, the intra-block dynamic ranges A and B required for requantization for each block are significantly reduced. I understand that. For this reason, the number of bits required for requantization can be significantly reduced.
[0053]
Here, when ADRC is used to compress, for example, 8-bit data into 3-bit data, the dynamic range in the block is DR, the bit allocation is p, the data level of the pixels in the block is x, and the requantization code is Q. As shown in FIG. 9A, requantization is performed by equally dividing the maximum value MAX and the minimum value MIN in the block with a designated bit length by the following equation (10).
[0054]
DR = MAX-MIN + 1
Q = [(x−MIN + 0.5) × 2 ^p / DR] (10)
[Equation 9]

[0055]
However,
[Expression 10]

Means the average of x satisfying Q = i, and [z] represents the largest integer equal to or smaller than z.
[0056]
Next, as shown in FIG. 9A, the highest gradation level (2) is obtained by p-bit requantization (in this case, 3-bit requantization). ^p The average value of the pixels in the block existing in the data level corresponding to -1) is taken, and this is set as the maximum value MAX ′ as shown in FIG. 9B. Further, as shown in FIG. 9A, the average value of the pixels in the block existing in the data level corresponding to the lowest gradation level 0 by this requantization is taken, and this is the minimum value MIN as shown in FIG. 9B. ′.
[0057]
Then, the in-block dynamic range DR ′ is newly redefined from the newly determined maximum value MAX ′ and the minimum value MIN ′, and the re-quantization code is defined as q, and the newly determined maximum value MAX ′ in the block and Based on the minimum value MIN ′, requantization is performed as shown in FIG.
[0058]
DR ′ = MAX′−MIN ′
q = [(x−MIN ′) × (2 ^p -1) /DR'+0.5] ( 11 )
However, [z] represents the maximum integer below z.
[0059]
In this way, by performing double requantization using ADRC, the dynamic range in the block is defined as a local feature of the image, and redundancy in the level direction can be mainly removed adaptively. Thus, efficient information compression is performed without being adversely affected by noise.
[0060]
Next, learning data subjected to compression processing using ADRC is divided into classes for each learning data. That is, the target pixel x shown in FIG. _Four 8 bit data x ₀ ~ X ₈ Q is the re-quantized data obtained by compressing the data into p bits by applying ADRC. ₀ ~ Q ₈ Then, the class (c) of the learning data is calculated using the following equation (12).
[0061]
[Expression 11]

[0062]
Here, FIG. 10 shows a flowchart as an example of a prediction value learning method based on the barycentric method. In the initialization of step S11, as preparation for performing this learning, 0 data is written to the class data table E (*) and the class frequency counter N (*). Here, “*” indicates all classes, the data table is E (C0), and the frequency counter corresponding to the class C0 is N (C0). When the control in step S11 ends, the control moves to step S12.
[0063]
In the class detection of step S12, class C is determined from the neighborhood data of the learning target pixel. For example, when 1-bit ADRC is applied to 8 neighboring pixels including the target pixel as in the above-described example, it is classified into 128 classes. As a method of class classification, in addition to ADRC as described above, classification methods such as PCM expression, DPCM, BTC, VQ, DCT, Hadamard transform, and the like can be considered. In addition, when considering the activity of a block made up of class classification target data, it is also possible to increase the number of classes by the type of classification by activity.
[0064]
Next, in the data detection in step S13, a target teacher signal e is detected. In the class-by-class data addition in step S14, the teacher signal e is added for each class C, and in the class-by-class frequency addition in step S15, the class C learning data frequency counter N (C) is incremented by +1. In step S16 for determining whether or not the control from step S12 to step S15 has been repeated for all the learning target data, if the learning of all the data has been completed, the control moves to step S17, and the learning target of all the data is determined. If not completed, the control moves to step S12. That is, in step S16, the control from step S12 to step S15 is repeatedly executed until learning of all data is completed, and the frequency table N (*) for all classes and the data table E (for all classes) ( *) Is generated.
[0065]
In the average value calculation for each class in step S17, the data integrated value that is the content of the data table E (*) of each class is divided by the frequency of the frequency counter N (*) of the corresponding class, and the average value of each class is calculated. Is calculated. This process is equivalent to calculating the center of gravity of the teacher signal distribution. This average value is the optimum predicted value of each class by the centroid method. In the average value registration for each class in step S18, learning by the center-of-gravity method is completed by registering the optimum predicted value corresponding to each class in a storage means such as a ROM, that is, this flowchart ends. As described above, in order to eliminate the influence of noise in the learning process, it may be considered that the case where the activity is small is excluded from the learning target. By the above-described prediction method and the center-of-gravity method, it is possible to predict a pixel value that is a more detailed position than a pixel.
[0066]
Next, an example of a learning method for the prediction coefficient ROM 30 used in the post-processing unit 23 will be described. First, as described above, an 8-bit signal is generated from a 10-bit teacher signal, and a prediction coefficient is obtained from the generated 8-bit signal and 10-bit teacher signal. An arbitrary 8-bit signal is converted into a 10-bit signal using the prediction coefficient. Then, the converted 10-bit signal is compared with the corresponding 10-bit teacher signal. As a result of the comparison, the 10-bit signal is discriminated into a value equal to the 10-bit teacher signal and a different value. At this time, the prediction coefficient is obtained as described above using the 10-bit teacher signal determined to have a different value and the converted 10-bit signal. The prediction coefficient obtained in this way is used for the prediction coefficient ROM 30.
[0067]
Further, the same method is used as the learning method of the optimum predicted value ROM 49 used in the post-processing unit 43. Specifically, the converted 10-bit predicted value is compared with the 10-bit teacher signal, and the 10-bit predicted value determined to be different from the 10-bit teacher signal is determined to be a different value. Using the 10-bit teacher signal and the converted 10-bit predicted value, an optimal predicted value is obtained as described above. The predicted value thus obtained is used in the optimal predicted value ROM 49.
[0068]
In these embodiments, the pixels used for classification and the pixels used for prediction calculation may be the same or different pixels.
[0069]
In the second embodiment, the prediction coefficient ROMs used in the class classification adaptive processing unit and the post-processing unit may be the same or different, for example, the number of prediction coefficients may be different.
[0070]
In the third embodiment, the optimum predicted value ROM used in the class classification adaptive processing unit and the post-processing unit may be the same or different.
[0071]
【The invention's effect】
According to the present invention, since the class classification adaptive processing is used, the amount of information corresponding to the increase in the number of quantization bits is not limited to simply increasing the number of quantization bits in the pixel data supplied from the outside. The signal can be converted into an increased signal, and generation of a pseudo contour due to insufficient gradation can be prevented.
[0072]
In addition, according to the present invention, since the class classification adaptive processing is used, it is easy to generate a digital signal having a larger number of bits than the input digital signal, and quantization noise of the generated digital signal is reduced. Reduced. As a result, image quality degradation due to the image process can be suppressed.
[Brief description of the drawings]
FIG. 1 is a block diagram used for explaining a prediction method applied to the present invention.
FIG. 2 is a schematic diagram showing the position of a pixel applied to the present invention.
FIG. 3 is a block diagram used for explaining a centroid method applied to the present invention.
FIG. 4 is a first embodiment of a quantization bit number conversion apparatus to which the present invention is applied.
FIG. 5 is a second embodiment of a quantization bit number conversion apparatus to which the present invention is applied.
FIG. 6 is a third embodiment of a quantization bit number conversion apparatus to which the present invention is applied.
FIG. 7 is an example of a prediction method learning method applied to the present invention;
FIG. 8 is a schematic diagram used for explaining ADRC applied to the present invention;
FIG. 9 is a schematic diagram used for explaining ADRC applied to the present invention;
FIG. 10 is an example of a learning method of the center of gravity method applied to the present invention.
FIG. 11 is a schematic diagram used to describe a quantization bit number conversion apparatus.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 12 ... Blocking circuit, 13 ... Prediction type arithmetic circuit, 14 ... ADRC circuit, 15 ... Class code generation circuit, 16 ... Prediction coefficient memory, 17 ... Level control circuit

Claims (7)

A level distribution pattern of the input digital signal is detected for each block composed of a plurality of pixels of the input digital signal having a plurality of pixels, and the block of the input digital signal belongs based on the detected pattern. A classifying means for determining a class and outputting class information;
Conversion information for converting into a digital signal quantized with a larger number of quantization bits than the input digital signal is stored for each class, and the conversion information is determined according to the class information from the class classification means. Storage means for outputting
Conversion means for converting the input digital signal into the digital signal based on the output conversion information;
When it is determined whether the quantized value of the digital signal is included in the existing section indicated by the quantized value of the input digital signal, and it is determined that the quantized value of the digital signal deviates from the existing section, A quantizing bit number conversion apparatus, comprising: limiting means for clipping the quantized value of the digital signal so that the quantized value of the digital signal falls within the existence section. A level distribution pattern of the first digital signal is detected for each block composed of a plurality of pixels of an input digital signal having a plurality of pixels, and the first digital signal is detected based on the detected pattern. First class classification means for determining the first class to which the block belongs and outputting first class information;
First conversion information for converting to a second digital signal quantized with a larger number of quantization bits than the first digital signal is stored for each of the first classes, and First storage means for outputting the first conversion information according to the first class information from the class classification means;
First conversion means for converting the first digital signal into a second digital signal based on the output first conversion information;
A level distribution pattern of the second digital signal is detected for each block of the converted second digital signal, and a second digital signal block to which the second digital signal block belongs is detected based on the detected pattern. Second class classification means for determining a class of the second class information and outputting second class information;
Second conversion information for converting the second digital signal into a third digital signal closer to the true value is stored for each of the second classes, and the second class classification means receives the above-mentioned information from the second class classification means. Second storage means for outputting the second conversion information according to second class information;
Second conversion means for converting the second digital signal into a third digital signal based on the output second conversion information;
Determining means for determining whether or not the existence value indicated by the quantized value of the first digital signal includes the quantized value of the second digital signal to be converted by the first converting means;
Based on the determination result, when the quantized value of the second digital signal obtained from the first conversion means is included in the existence section, the second digital signal obtained from the first conversion means And selecting the third digital signal obtained from the second conversion means when the quantized value of the second digital signal obtained from the first conversion means deviates from the existence section Means for converting the number of quantized bits. In the quantization bit number conversion device according to claim 2,
In the first conversion means,
A linear first-order combination of the first digital signal and the first conversion information is converted into the second digital signal;
In the second conversion means,
A quantizing bit number conversion apparatus, wherein the linear digital combination of the second digital signal and the second conversion information is converted into the third digital signal. A level distribution pattern of the first digital signal is detected for each block composed of a plurality of pixels of an input digital signal having a plurality of pixels, and the first digital signal is detected based on the detected pattern. First class classification means for determining the first class to which the block belongs and outputting first class information;
A second digital signal quantized with a larger number of quantization bits than the first digital signal is stored for each first class, and the first class from the first class classification means is stored. First storage means for outputting the second digital signal in response to information;
A level distribution pattern of the second digital signal is detected for each block of the output second digital signal, and a second digital signal block to which the second digital signal block belongs is detected based on the detected pattern. Second class classification means for determining a class of the second class information and outputting second class information;
A third digital signal closer to the true value than the second digital signal is stored for each second class, and the third digital signal is stored in accordance with the second class information from the second class classification means. Second storage means for outputting a digital signal of
Determination means for determining whether or not the existence value indicated by the quantization value of the first digital signal includes the quantization value of the second digital signal obtained from the first storage means;
When the quantization value of the second digital signal obtained from the first storage means is included in the existence section based on the determination result, the second digital signal obtained from the first storage means And when the quantized value of the second digital signal obtained from the first storage means deviates from the existing section, the third digital signal obtained from the second storage means is selected. And a quantizing bit number converting device. A level distribution pattern of the input digital signal is detected for each block composed of a plurality of pixels of the input digital signal having a plurality of pixels, and the block of the input digital signal belongs based on the detected pattern. A class classification step for determining a class and outputting class information;
Conversion information for converting into a digital signal quantized with a larger number of quantization bits than the input digital signal is stored in the storage means for each class, and according to the class information from the class classification step Outputting the conversion information from the storage means;
A conversion step of converting the input digital signal into the digital signal based on the output conversion information;
When it is determined whether the quantized value of the digital signal is included in the existing section indicated by the quantized value of the input digital signal, and it is determined that the quantized value of the digital signal deviates from the existing section, And a limiting step of clipping the quantized value of the digital signal so that the quantized value of the digital signal falls within the existing section. A level distribution pattern of the first digital signal is detected for each block composed of a plurality of pixels of an input digital signal having a plurality of pixels, and the first digital signal is detected based on the detected pattern. A first class classification step of determining a first class to which the block belongs and outputting first class information;
First conversion information for converting to a second digital signal quantized with a larger number of quantization bits than the first digital signal is stored in the first storage means for each first class. Outputting the first conversion information from the first storage means in response to the first class information from the first class classification step;
A first conversion step of converting the first digital signal into a second digital signal based on the output first conversion information;
A level distribution pattern of the second digital signal is detected for each block of the converted second digital signal, and a second digital signal block to which the second digital signal block belongs is detected based on the detected pattern. A second class classification step for determining a class of the second class and outputting second class information;
Second conversion information for converting the second digital signal into a third digital signal closer to a true value is stored in the second storage means for each of the second classes, and the second Outputting the second conversion information from the second storage means according to the second class information from the class classification step;
A second conversion step of converting the second digital signal into a third digital signal based on the output second conversion information;
A determination step of determining whether or not the existence section indicated by the quantization value of the first digital signal includes the quantization value of the second digital signal converted in the first conversion step;
When the quantized value of the second digital signal obtained from the first conversion step is included in the existence section based on the determination result, the second digital signal obtained from the first conversion step. And selecting the third digital signal obtained from the second conversion step when the quantized value of the second digital signal obtained from the first conversion step deviates from the existence section. A method for converting the number of quantized bits. A level distribution pattern of the first digital signal is detected for each block composed of a plurality of pixels of an input digital signal having a plurality of pixels, and the first digital signal is detected based on the detected pattern. A first class classification step of determining a first class to which the block belongs and outputting first class information;
A second digital signal quantized with a larger number of quantization bits than the first digital signal is stored in the first storage means for each first class, and from the first class classification step. Outputting the second digital signal from the first storage means according to the first class information of:
A level distribution pattern of the second digital signal is detected for each block of the output second digital signal, and a second digital signal block to which the second digital signal block belongs is detected based on the detected pattern. A second class classification step for determining a class of the second class and outputting second class information;
A third digital signal closer to the true value than the second digital signal is stored in the second storage means for each second class, and the second class information from the second class classification step is stored. In response to outputting the third digital signal from the second storage means;
A determination step for determining whether or not the existence section indicated by the quantization value of the first digital signal includes the quantization value of the second digital signal obtained from the first storage means;
When the quantization value of the second digital signal obtained from the first storage means is included in the existence section based on the determination result, the second digital signal obtained from the first storage means And when the quantized value of the second digital signal obtained from the first storage means deviates from the existing section, the third digital signal obtained from the second storage means is selected. A method for converting the number of quantized bits.

Images