Deep metric learning via subtype fuzzy clustering

Abstract

Deep metric learning methods aim to transform data features from original scattered space to a discriminative subspace in an end-to-end way, and they have shown promising results on wide applications. Triplet loss functions are the most popular models to tackle deep metric learning problem as they simultaneously enhance separability between different classes and compactness of each class in the embedded subspace. Therefore, effective triplets selection is crucial to the classification performance. However, most of these methods only focus on mining hard negative pairs, which refer to the nearest sample pairs of different classes, while fail to take subtle cluster structure of each class into consideration. To take such information into the metric learning model, a novel scheme based on subtype fuzzy clustering is proposed in this paper. By defining a new clustering degree, we conduct fuzzy clustering for each class to mine classification-oriented subtype structure. The new clustering degree is inversely proportional to the pairwise distance, thus, we can choose the positive pairs of the highest clustering degrees directly based on the farthest distances within each class. This new sampling approach avoids off-line clustering step, for which the network weights update procedure has to be temporarily paused. In other words, our method builds positive pairs without explicit clustering degree computation or off-line clustering. Our method inputs the selected positive pairs and negative pairs into the standard triplet loss to achieve network feature learning. Experimental results show competitive metric learning performance on three benchmark datasets.