Unsupervised Detection of Outlier Images Using Multi-Order Image Transforms

Abstract

The task of unsupervised detection of peculiar images has immediate applications to numerous scientific disciplines
such as astronomy and biology. Here we describe a simple non-parametric method that uses multi-order image
transforms for the purpose of automatic unsupervised detection of peculiar images in image datasets. The method is
based on computing a large set of image features from the raw pixels and the first and second order of several combinations
of image transforms. Then, the features are assigned weights based on their variance, and the peculiarity of
each image is determined by its weighted Euclidean distance from the centroid such that the weights are computed
from the variance. Experimental results show that features extracted from multi-order image transforms can be used
to automatically detect peculiar images in an unsupervised fashion in dierent image datasets, including faces, paintings,
microscopy images, and more, and can be used to find uncommon or peculiar images in large datasets in cases
where the target image of interest is not known. The performance of the method is superior to general methods such
as one-class SVM. Source code and data used in this paper are publicly available, and can be used as a benchmark to
develop and compare the performance of algorithms for unsupervised detection of peculiar images.