文献类型：期刊文献

英文题名：BDLA: Bi-directional local alignment for few-shot learning

作者：Zheng, Zijun[1,2];Feng, Xiang[1,2];Yu, Huiqun[1,2];Li, Xiuquan[3];Gao, Mengqi[1,2]

机构：[1]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Shanghai Engn Res Ctr Smart Energy, Shanghai 200237, Peoples R China;[3]Chinese Acad Sci & Technol Dev, Beijing 100038, Peoples R China

年份：2023

卷号：53

期号：1

起止页码：769

外文期刊名：APPLIED INTELLIGENCE

收录：;EI(收录号：20221712033706);WOS:【SCI-EXPANDED(收录号:WOS:000784720600002)】；

基金：This work was supported in part by the Key Program of the National Natural Science Foundation of China under Grant No. 62136003, the National Natural Science Foundation of China under Grant Nos. 61772200 and 61772201, Shanghai Pujiang Talent Program under Grant No. 17PJ1401900, Shanghai Economic and Information Commission "Special Fund for Information Development" under Grant No. XX-XXFZ-02-20-2463.

语种：英文

外文关键词：Few-shot learning; Local descriptor; Bi-directional distance; Convex combination

摘要：Deep learning has been successfully exploited to various computer vision tasks, which depend on abundant annotations. The core goal of few-shot learning, in contrast, is to learn a classifier to recognize new classes from only a few labeled examples that produce a key challenge of visual recognition. However, most of the existing methods often adopt image-level features or local monodirectional manner-based similarity measures, which suffer from the interference of non-dominant objects. To tackle this limitation, we propose a Bi-Directional Local Alignment (BDLA) approach for the few-shot visual classification problem. Specifically, building upon the episodic learning mechanism, we first adopt a shared embedding network to encode the 3D tensor features with semantic information, which can effectively describe the spatial geometric representation of the image. Afterwards, we construct a forward and a backward distance by exploring the nearest neighbor search to determine the semantic region-wise feature corresponding to each local descriptor of query sets and support sets. The bi-directional distance can encourage the alignment between similar semantic information while filtering out the interference information. Finally, we design a convex combination to merge the bi-directional distance and optimize the network in an end-to-end manner. Extensive experiments also show that our proposed approach outperforms several previous methods on four standard few-shot classification datasets.