MULTI-SCALE TEXTURE FEATURE REPRESENTATION FOR CONTENT-BASEDIMAGE DATABASE RETRIEVAL

摘要

Recent systems for content-based image retrievalrn(CBIR) employ multi-scale image filtering techniquesrnsuitable for texture analysis. Although they comparernfavourably to alternative techniques that do notrnemploy convolution filters, these multi-scale CBIRrnsystems perform rather poorly when compared tornhuman observers. This seems not only due to thernpeculiar ability of humans to infer visual featuresrnand semantic meanings from images based on priorrnknowledge, but also to the similarity inaccuracy introducedrnby: I) the feature representation (I.e., thernimage characteristic signature extraction), which isrnintrinsically non-injective, and ii) the similarity measure,rnwhose selection depends on the set of features.rnThis work reports on new developments in featurernextraction, feature representation and distance measurernselection for content- based image retrieval.