Nearest Neighbor with Multi-feature Metric for Image Annotation

Abstract

Most of the Nearest Neighbor (NN) based image annotation (or classification) methods cannot achieve satisfactory performance. In this paper, we propose a novel Nearest Neighbor method based on a multi-feature distance metric, which takes full advantage of different and complementary features. We first establish a metric for each feature and assign a weight for every metric, and then linearly combine all of them together to form one distance metric, namely the multi-feature metric. After that, we construct an NN model based on “image-to-cluster” distances, which equals to the distances between an image and the clusters within an image category using our multi-feature based metric, and which is different from calculating Euclidean distances between two images. By introducing this multi-feature based distance metric, our NN based model can mitigate the semantic issues due to intra-class variations and inter-class similarities, and improve the image annotation performance. Experiments confirm the superiority of our model in comparison with both the traditional classifiers and the state of the art learning-based models.