公开(公告)号：US11205098B1

公开(公告)日：2021-12-21

申请号：US17373831

申请日：2021-07-13

Applicant: INSTITUTE OF AUTOMATION, CHINESE ACADEMY OF SCIENCES

Inventor： Zhengxing Wu ,  Junzhi Yu ,  Yue Lu ,  Xingyu Chen

IPC: G06K9/00 ,  G06K9/62 ,  G06T7/70 ,  G06T7/60 ,  G06N3/04 ,  G06N3/08 ,  G06K9/46

Abstract: A single-stage small-sample-object detection method based on decoupled metric is provided to solve the following problems: low detection accuracy of existing small-sample-object detection methods, the mutual interference between classification and regression in a non-decoupled form, and over-fitting during training of a detection network in case of small samples. The method includes: obtaining a to-be-detected image as an input image; and obtaining a class and a regression box corresponding to each to-be-detected object in the input image through a pre-constructed small-sample-object detection network DMNet, where the DMNet includes a multi-scale feature extraction network, a decoupled representation transformation module, an image-level distance metric learning module and a regression box prediction module. The new method avoids the over-fitting during training of the detection network, eliminates the mutual interference between the classification branch and the regression branch, and improves the accuracy of small-sample-object detection.