上下文感知顯著性檢測(cè)

Context-AwareSaliencyDetection

Anewtypeofsaliency–context-awaresaliency1.Introduction2.Abstractwepresentadetectionalgorithmwhichisbasedonfourprinciplesobservedinthepsychologicalliteratureevaluatedintwoapplications:imageretargetingandsummarization

2.Principlesofcontext-awaresaliencyfourprinciples

Locallow-levelconsiderations,includingfactorssuchascontrastandcolor.

Globalconsiderations,whichsuppressfrequently-occurringfeatures,whilemaintainingfeaturesthatdeviatefromthenorm.

Visualorganizationrules,whichstatethatvisualformsmaypossessoneorseveralcentersofgravityaboutwhichtheformisorganized.High-levelfactors,suchashumanfaces.

2.Principlesofcontext-awaresaliency(b)沒(méi)有抑制常見(jiàn)特征(c)未考慮局部對(duì)比，抑制太多特征saliency區(qū)域太少(d)兩者結(jié)合(e)還檢測(cè)到了水珠3.Detectionofcontext-awaresaliencyLocal-globalsingle-scalesaliency:當(dāng)以pixel

i為中心的patch

i和圖片上很多patch

j的距離都很大時(shí)，認(rèn)為pixel

i是saliency的

dcolor:Euclideandistancebetweenthevectorizedpatchespi

andpj

inCIEL*a*bcolorspace——Principle(1)~(2)dposition:Euclideandistancebetweenthevectorizedpatchespi

andpj

inCIEL*a*bcolorspace——Principle(3)3.Detectionofcontext-awaresaliencyLocal-globalsingle-scalesaliency:K=65；r:某種尺度取跟patch

i距離最近的65個(gè)patch

q，不需要比較所有的patch，因?yàn)槿绻罱膒atch跟patch

i的距離很大，則說(shuō)明patch

i很顯著，顯著性公式（2）為單尺度下的顯著性值：3.Detectionofcontext-awaresaliencyMulti-scalesaliencyenhancement

rk:多種尺度尺度選擇：{100%,80%,50%,30%}計(jì)算多個(gè)尺度下的顯著性值，取平均值

3.Detectionofcontext-awaresaliencyIncludingtheimmediatecontext:修正之前得到的顯著性區(qū)域：根據(jù)閾值，顯著性值大于0.8的當(dāng)作mostattendedlocalizedareas圖片上其他的像素，根據(jù)公式，重新計(jì)算顯著性值dfoci(i)為像素i與距離最近的，屬于mostattendedlocalizedareas的像素的歐式距離3.Results

對(duì)應(yīng)figure

1的三種情況3.Results

對(duì)應(yīng)figure

1的三種情況3.Results

對(duì)應(yīng)figure

1的三種情況3.Results

量化結(jié)果4.Application

Imageretargeting4.Application

Imageretargeting4.Application

summarization5.Conclusion

提出一種新的saliency；提出四張標(biāo)準(zhǔn)局部對(duì)比（顏色、圖案），具有不同顏色和圖