Search International and National Patent Collections

1. (WO2018137358) DEEP METRIC LEARNING-BASED ACCURATE TARGET RETRIEVAL METHOD

Pub. No.:    WO/2018/137358    International Application No.:    PCT/CN2017/104397
Publication Date: Fri Aug 03 01:59:59 CEST 2018 International Filing Date: Sat Sep 30 01:59:59 CEST 2017
IPC: G06F 17/30
Applicants: PEKING UNIVERSITY
北京大学
Inventors: DUAN, Lingyu
段凌宇
BAI, Yan
白燕
LOU, Yihang
楼燚航
GAO, Feng
高峰
Title: DEEP METRIC LEARNING-BASED ACCURATE TARGET RETRIEVAL METHOD
Abstract:
Deep metric learning-based accurate target retrieval method, the method comprising: in the iterative training of a deep neural network structure, enabling the same class of target objects to be close to each other and enabling different classes of target objects to be far away from each other in a process wherein characteristics of multiple extracted pictures of the same class of target object are processed, wherein intra-class individuals having a similar attribute are close to each other in the characteristic distribution of individuals of the same class, and the distance between intra-class individuals having different attributes is greater than a preset distance, so as to obtain a trained deep neural network structure; using a trained deep neural network structure to extract respective characteristics from pictures to be queried and a preset reference picture respectively; and obtaining Euclidean distances of characteristics between the query pictures and the reference picture, and sorting the distances from small to large so as to obtain an accurate retrieval target. With said method, the problem of accurate retrieval in a vertical field is solved.