Ranking-based triplet loss function with intra-class mean and variance for fine-grained classification tasks
- PDF / 2,489,351 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 23 Downloads / 175 Views
METHODOLOGIES AND APPLICATION
Ranking-based triplet loss function with intra-class mean and variance for fine-grained classification tasks J. Bhattacharya1 · R. K. Sharma1
© Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract This paper proposed a deep ranking model for triplet selection to efficiently learn similarity metric from top ranked images. A modified distance criterion described in the current work leverages the intra-category variance in metric learning of a triplet network by learning a local sample structure. A multicolumn fusion architecture is used to capture different levels of variance, which when incorporated in the loss function strengthens it and optimizes the objective of the triplet networks. This enables a fine-grained classification strategy. State-of-the-art techniques use a group-sensitive triplet sampling to deal with this issue. However, these have the disadvantage of increased group sampling computations. Experiments are conducted over a variety of benchmark datasets including Model40, PatternNet, and In-Shop Clothing. The main purpose of these experiments are to verify whether the triplet learning technique can be applied over different kinds of data. Results demonstrate that the current work provides superior results in most cases. These results can further be improved with specific parameter tunings and ensembling techniques wherever applicable. Keywords Image classification · Triplet learning · Convolutional neural network
1 Introduction Availability of large labeled datasets consisting of billions of images, and computationally powerful resources have motivated the design of deep convolutional networks for fine-grain visual recognition. These models are efficient in providing a data representation which can encapsulate subtle feature information invariant to structural and illuminational conditions. Nevertheless, inter-class similarity and intraclass variance of the feature embeddings pose a challenge for verification tasks which are an extension of the original classification network and hence the performance lack is considered as a side effect caused by the metric learnt during classification. Literature reports that this problem has been dealt with in two ways. In the first approach, a metric is learnt from an intermediate bottleneck layer data representation of Communicated by V. Loia.
B
J. Bhattacharya [email protected] R. K. Sharma [email protected]
1
Thapar Institute of Engineering and Technology, CSED, Patiala 147004, India
the network while feature embedding itself is optimized in the second approach. Some specific contributions of the first approach include large margin nearest neighbor (LMNN) metric, Mahalanobis distance learning using the information theoretic metric learning (ITML) method, joint Bayesian approach, weighted χ 2 similarity on the nonnegative sparse normalized DeepFace representations. An example of learning feature embeddings is training a Siamese network using standard cross-entropy loss to learn the feature pair distance. In
Data Loading...
