Title:
IMAGE QUESTION ANSWERING METHOD, DEVICE AND SYSTEM, AND STORAGE MEDIUM
Document Type and Number:
WIPO Patent Application WO/2019/085905
Kind Code:
A1
Abstract:
The present application provides an image question answering method, device and system, and a storage medium. The image question answering method comprises: extracting a question feature representing semantics of a question, a global feature of an image, and a detection box feature of a detection box surrounding an object in the image; obtaining, on the basis of the question feature, the global feature, and the detection box feature, a first weight of each of at least one region of the image and a second weight of each of at least one detection box of the image; performing weighing processing on the global feature using the first weight to obtain a regional attention feature of the image; performing weighing processing on the detection box using the second weight to obtain a detection box attention feature of the image; and predicting an answer to the question on the basis of the question feature, the regional attention feature, and the detection box attention feature. The technical solution provided by the present application can comprehensively take a global feature and a detection box feature comprising an object detection box of an image into consideration, thereby improving the efficiency and precision of image question answering.
Inventors:
LU PAN (CN)
LI HONGSHENG (CN)
WANG XIAOGANG (CN)
LI HONGSHENG (CN)
WANG XIAOGANG (CN)
Application Number:
PCT/CN2018/112735
Publication Date:
May 09, 2019
Filing Date:
October 30, 2018
Export Citation:
Assignee:
BEIJING SENSETIME TECH DEVELOPMENT CO LTD (CN)
International Classes:
G06F16/50
Foreign References:
| CN108228703A | 2018-06-29 | |||
| CN106649542A | 2017-05-10 | |||
| CN105574133A | 2016-05-11 | |||
| CN107256228A | 2017-10-17 | |||
| US20120290577A1 | 2012-11-15 |
Attorney, Agent or Firm:
BEIJING SIYUANZHIHUI INTELLECTUAL PROPERTY AGENCY (CN)
Download PDF: