Adaptive Motion-compensated Temporal Noise Reduction

Motion-compensation就是detect出motion region出來,不做3DNR以避免出現殘影。為了確保影像裡的motion region不會被3DNR給抹掉,所以在式子裡的w會依據{ Image }_{ i }{ Base }_{ i-1 }的pixel差值來決定。差值越大,Base的blending weight值就要越小。為了確保計算motion region時不會被random noise給干擾,在計算差值時,基本上不會用點對點相減,而是用Block-Based的Motion Detector來減少random nosie造成的干擾。

{ Base }_{ i }=({ w }_{ i })*{ Base }_{ i-1 }+(1-{ w }_{ i })*{ Image }_{ i }

{ Diff }_{ i }(x,y)={ abs(CurrentFrame }_{ i }(x,y)-{ BaseFrame }_{ i-1 }(x,y))

 

3DNR[Calculate the Different in Bayer Pattern.]

之後再把Different轉換成Stable Level做為3DNR Algorithm裡的Weight(或上面式子裡w)大小的依據,如此一來,可以避免motion region不會被3DNR給抹除。Stable Level影像裡越白的地方就是motion region。

3DNR                                           [Transfer Different to Stable Level]

結果如下所示(下圖是經過demosaic之後的影像,這裡的3DNR是做在Bayer Pattern裡),喝水鳥的頭因為隨著時間在移動,被detect為motion region不會被3DNR給抹除,而其他靜態的場景依然能保有3DNR過後較好的影像品質。3DNR[Result Compare]

 

 


reference : 

Adaptive motion compensated filtering of noisy image sequences(1993)

Motion-compensated temporal filtering and motion vector coding using biorthogonal filters(2005)

Leave a Reply

Your email address will not be published. Required fields are marked *