کنفرانس بین المللی فناوری اطلاعات و دانش

English

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Intra Class Feature Learning and Supervised Triplet Sampling for Deep Metric Learning

نویسندگان :

Hamideh Rafiee¹ Ahmad Ali Abin² Seyed Soroush Majd³ Viet-Vu Vu⁴

1- shahid beheshti university 2- shahid beheshti university 3- shahid beheshti university 4- CMC University

کلمات کلیدی :

deep metric learning،zero-shot learning،similarity learning

چکیده :

Deep metric learning (DML) aims to learn an embedding space where semantically similar samples are mapped close together while dissimilar ones are placed farther apart. Although effective on seen classes, most DML approaches mainly emphasize inter-class separation during training and consequently struggle to generalize to unseen classes during inference. To address this limitation, we propose a method that improves generalization by learning fine-grained intra-class features and employing a supervised triplet sampling strategy. The first component, fine-grained feature learning, prevents over-compression in the embedding space by capturing structural relationships among samples within the same class, allowing the model to represent subtle intra-class variations alongside class-discriminative features. The second component, supervised triplet sampling, selects informative anchor, positive, and negative samples according to a discrimination uncertainty criterion derived from the classifier’s prediction ambiguity, ensuring that challenging examples contribute effectively to the optimization process. Experiments conducted on the Cars196 dataset demonstrate that the proposed approach achieves notable improvements in both clustering and retrieval tasks.