A Preliminary Study of Spatial Bias in Knn Distance Metrics

A machine learning algorithm for image classification exhibits spatial bias if permuting the order of image pixels significantly alters its classification accuracy. In this paper, we explore the spatial bias of a number of different distance metrics for k-nearest-neighbor image classification. One distance metric is inspired by the convolutional kernels employed in convolutional neural networks. The other metrics are based on BRIEF descriptors, which generate bit vectors corresponding to images based on comparisons of pixel intensity values. We found that the convolutional distance metric exhibited a strong positive spatial bias, as did one of the BRIEF descriptors. Another BRIEF descriptor exhibited a negative spatial bias, and the remainder exhibited little or no spatial bias. These results lay a foundation for future work that would involve larger numbers of convolutional iterations, potentially synergized with BRIEF-style image preprocessing.