1、基于哈希的图像搜索技术提纲提纲背景典型方法评价指标总结 背景图片搜索引擎图片搜索引擎移动搜索业移动搜索业务务皮皮革革纺纺织织医医学学背景背景为什么要用哈希？背景背景怎样对图像特征进行哈希编码？背景背景 背景11011001011000100000000112()(),(),.,()0,1KiiiKiH xh xh xhx 背景11011001011000100000000112()(),(),.,()0,1KiiiKiH xh xh xhxHamming distance:14 背景 背景100001110111000101Dataset()H XSimple hash table 背景100
2、001110111000101Dataset()H XQ()qH x110111Simple hash tableXOR 背景100001110111000101 NQDatasetSearch the hash table for a small set of images.()H XQ()qH x110111Simple hash tableXOR 背景背景背景特征进行哈希编码的本质？背景背景Gionis,A.&Indyk,P.&Motwani,R.(1999)Take random projections of dataQuantize each projection with few
3、bits010101101No learning involvedGist descriptor 典型方法 局部局部敏感哈希敏感哈希 Local Sensitive HashingThe probability that a random hyperplane separates two unit vectors depends on the angle between them:Goemans and Williamson 1995,Charikar 2004,modified by meHigh dot product:unlikely to splitLower dot product:
4、likely to splitCorresponding hash function:典型方法 局部局部敏感哈希敏感哈希 Local Sensitive HashingGround Truth32 bit128 bitA LSH family,H(c,r,P1,P2),for any p,q belong to S,If|p-q|r,then Pr(h(p)=h(q)P1 If|p-q|cr,then Pr(h(p)=h(q)P2256 bit 典型方法 局部局部敏感哈希敏感哈希 Local Sensitive Hashing 典型方法 局部局部敏感哈希敏感哈希 Local Sensitive
5、 Hashing 典型方法 迭代量化编码迭代量化编码 Iterative QuantizationAverage quantization error:1.00(a)PCA aligned.Average quantization error:0.93(b)Random Rotation.Average quantization error:0.88(c)Optimized Rotation.Q(B,R)=|B-VR|F21000011111100001010010111.S.Lazebnik,A.Gordo,and F.Perronnin,“Iterative Quantization:A
6、Procrustean Approach to Learning Binary Codes for Large-scale Image Retrieval,”Accepted,IEEE Trans.Pattern Analysis and Machine Learning Intelligence,2012.典型方法 迭代量化编码迭代量化编码 Iterative Quantization1.S.Lazebnik,A.Gordo,and F.Perronnin,“Iterative Quantization:A Procrustean Approach to Learning Binary Co
7、des for Large-scale Image Retrieval,”Accepted,IEEE Trans.Pattern Analysis and Machine Learning Intelligence,2012.典型方法 迭代量化编码迭代量化编码 Iterative QuantizationRRITQ32 bitLSHPCAH 典型方法 迭代量化编码迭代量化编码 Iterative Quantization 典型方法迭迭稀疏迭代量化编码稀疏迭代量化编码 Sparse Projections for High-Dimensional Binary Codes1.Y.Xia,K.He
8、,P.Kohli,and J.Sun,“Sparse Projections for High-Dimensional Binary Codes,”in CVPR,2015.典型方法 核监督哈希核监督哈希 Supervised Hashing with Kernels1.W.Liu,W.Jun,R.Ji,Y.Jiang,and S.Chang,“Supervised hashing with kernels,”Proc.IEEE Conf.Computer Vision and Pattern Recognition,pp.2074-2081,2013.structure topology based评价指标评价指标semantic topology based评价指标评价指标总结总结谢谢谢谢
