Combining multi-scale dissimilarities for image classification

摘要

In image classification, multi-scale information is usually combined by concatenating features or selecting scales. Their main drawbacks are that concatenation increases the feature dimensionality by the number of scales and scale selection typically loses the information from other scales. We propose to solve this problem by the dissimilarity representation as it enables to combine various sources of information without increasing the dimensionality of the representation space. Various combining rules are introduced and tested with real-world applications. Our experiments show that combining with dissimilarities from all scales could indeed improve considerably upon the performance of the best single scale and adaptive combining can improve upon straightforward averaging.