基于稀疏编码与反向索引的鞋印图像比对算法

Shoeprint Image Matching Algorithm Based on Sparse Coding and Invert Indexing

针对刑侦工作中大规模鞋印图像库的查询应用需求,提出一种基于稀疏编码与反向索引的快速比对算法。首先,对鞋印图像进行视觉增强、中值滤波与二值分割等预处理,并提取其尺度不变特征变换(SIFT)特征;然后,基于聚类字典构造、稀疏编码(SC)与最大池化等方法,计算鞋印图像的稀疏编码特征;最后,通过构建"词-图像矩阵"而建立每个视觉单词的反向索引(RI)表,并据此提出一种SC-RI鞋印图像比对新算法。基于16 343幅真实鞋印图像的试验结果表明,SC-RI算法完成一次比对平均耗时约为121.26ms,较之传统SIFT匹配穷举比对方法,其速度提高了140多倍,且局部花纹比对TOP 20正确率可达到95.3%。

Focused at the query application into the large-scale shoeprint image library for criminal investigation,a fast matching algorithm,based on sparse coding and invert indexing,was proposed here.Firstly,the shoeprint images were preprocessed by the manipulation of visional enhancement,median-value filter and binary segmentation so as to extract the Scale Invariable Feature Transformation(SIFT)specifics for every shoeprint image.Then,in combination of clustering dictionary learning,sparse coding(SC)and maximum pooling handling,a calculation was carried out on the sparse coding feature of shoeprint images.Finally,through a“word-image matrix”to be constructed,a reverse indexing(RI)table was set up for each visual word so that a new algorithm,named SC-RI,was set up for fast matching of shoeprint image.Experimental results from 16343 real shceprintsimages showed that the SC-RI algorithm fulfilled a matching within about 121.26 milliseconds,being 140 times higher than that of the traditional SIFT-exhaustive matching choice,making the TOP 20’s accuracy reach up to 95.3%for local pattern matching in shoeprint image query.