Combining Hiearachical Clustering and Naive-Bayes Nearest-Neighbor For Image Classification

Authors

Chen Fu, Shijie Jia

Corresponding Author

Chen Fu

DOI

https://doi.org/10.2991/lemcs-15.2015.7How to use a DOI?

Keywords

Naïve Nearst Neighbour Classifier; Image classification; Imageto-Class distance hierarchical clustering; Category cutting

Abstract

This paper addresses the problem of image classification, which has successful applications in many fields, such as image retrieve, object detection and recognition. As a simple nonparametric methods, Naive-Bayes Nearest-Neighbor(NBNN) employs NN distances in the space of the local image descriptors and computes direct ‘Image-to-Class’ distances without descriptor quantization. However, the high compute cost is the severe bottleneck for the generation of NBNN, which is particular for large scale image classification. Under the naïve bayes assumptation (the local features are considered i.i.d.), this paper combines hierarchical clustering with Naive-Bayes nearest-neighbor for image classification, which provides high scalability to the tradeoff between accuracy and efficiency. Furthermore, this paper proposes the scheme of category cutting, which could improve the test speed at the cost of little accuracy dropping, which is particularly efficient for large scale image classification. Experiments showed that our scheme is the most efficiency among LSH, kd-tree, and linear search algorithms.

Open Access

This is an open access article distributed under the CC BY-NC license.