Combining dissimilarity measures for image classification

•Proposing a novel method for image classification.•Combining global and local dissimilarity measures from two sample spaces for image classification.•Proposing three novel distance ratios for distance metric.•Revealing the difference between the Euclidean and block distances.•Extensive experimental results are conducted. The local dissimilarity has been verified as one of effective metrics for pattern classification. For high-dimensional data, because of the property of high-dimension, even if there are a number of available samples, they are still only resultant observations of a sampling process of the high-dimensional population. As a consequence, available samples at least partially possess the property of randomness and are not “accurate” representations of the true and total sample space. Besides the prevailing local dissimilarity measure, global dissimilarity measures might also be exploited for improving the classification approach. In this paper, we propose to directly exploit global and local dissimilarity measures to efficiently perform image classification. The proposed method proposes to simultaneously use three dissimilarities derived from the original and transform sample space. These dissimilarities including the elaborated distance ratio enable space relations of the probe sample and gallery samples to be measured from three viewpoints, so the combination of them provide us with more reliable measurements on spatial geometric relationship of samples. An obvious advantage of this combination is that we can attain a very robust evaluation on the space distance between samples and the consequent classification decision will be less affected by the noise in data. The experiments prove that the proposed method does achieve the desired goal, i.e., very satisfactory accuracy improvement in comparison with the previous state-of-the art methods. [Display omitted]