基于哈希的圖像搜索技術(shù)

基于哈希的圖像搜索技術(shù)第一期提綱背景典型方法評(píng)價(jià)指標(biāo)總結(jié)背景圖片搜索引擎移動(dòng)搜索業(yè)務(wù)皮革紡織醫(yī)學(xué)背景為什么要用哈希？背景怎樣對(duì)圖像特征進(jìn)行哈希編碼？背景…背景…110110010110001000000001…背景110110010110001000000001……Hammingdistance:14……背景背景100001110111000101DatasetSimplehashtable背景100001110111000101DatasetQ110111SimplehashtableXOR背景100001110111000101<<NQDatasetSearchthehashtableforasmallsetofimages.Q110111SimplehashtableXOR背景背景特征進(jìn)行哈希編碼的本質(zhì)？背景Gionis,A.&Indyk,P.&Motwani,R.(1999)TakerandomprojectionsofdataQuantizeeachprojectionwithfewbits010101101NolearninginvolvedGistdescriptor典型方法

局部敏感哈希

LocalSensitiveHashingTheprobabilitythatarandom

hyperplaneseparatestwounitvectorsdependsontheanglebetweenthem:[GoemansandWilliamson1995,Charikar2004,modifiedbyme]Highdotproduct:unlikelytosplitLowerdotproduct:likelytosplitCorrespondinghashfunction:典型方法

局部敏感哈希

LocalSensitiveHashingGroundTruth32bit128bitALSHfamily,H(c,r,P1,P2),foranyp,qbelongtoS,If||p-q||≤r,thenPr(h(p)=h(q))≥P1If||p-q||≥cr,thenPr(h(p)=h(q))≤P2256bit典型方法

局部敏感哈希

LocalSensitiveHashing典型方法

局部敏感哈希

LocalSensitiveHashing典型方法

迭代量化編碼

IterativeQuantizationAveragequantizationerror:1.00(a)PCAaligned.Averagequantizationerror:0.93(b)RandomRotation.Averagequantizationerror:0.88(c)OptimizedRotation.Q(B,R)=||B-VR||F2100001111110000101001011[1].S.

Lazebnik,A.Gordo,andF.Perronnin,“IterativeQuantization:AProcrusteanApproachtoLearningBinaryCodesforLarge-scaleImageRetrieval,”Accepted,IEEETrans.PatternAnalysisandMachineLearningIntelligence,2012.典型方法

迭代量化編碼

IterativeQuantization[1].S.

Lazebnik,A.Gordo,andF.Perronnin,“IterativeQuantization:AProcrusteanApproachtoLearningBinaryCodesforLarge-scaleImageRetrieval,”Accepted,IEEETrans.PatternAnalysisandMachineLearningIntelligence,2012.典型方法

迭代量化編碼

IterativeQuantizationRRITQ32bitLSHPCAH典型方法

迭代量化編碼

IterativeQuantization典型方法迭稀疏迭代量化編碼

SparseProjectionsforHigh-DimensionalBinaryCodes[1].Y.Xia,K.He,P.Kohli,andJ.

Sun,“SparseProjectionsforHigh-DimensionalBinaryCodes,”inCVPR,2015.典型方法

核監(jiān)督哈希

SupervisedHashingwithKernels[1].W.Liu,W.Jun,R.Ji,Y.Jiang,andS.Chang,“Supervisedhashingwithkernels,”Proc.IEEECo