Title:
IMAGE RETRIEVAL METHOD BASED ON DEEP LEARNING AND SEMANTIC SEGMENTATION
Document Type and Number:
WIPO Patent Application WO/2019/237646
Kind Code:
A1
Abstract:
An image retrieval method based on deep learning and semantic segmentation, comprising the following steps of: reading an image and performing preprocessing on the image; encoding the image into a set of feature images by any one convolution layer of a deep neural network by means of depth learning; performing semantic segmentation on the image and obtaining a pixel-by-pixel category label of the segmented image; performing weighting processing on the feature images according to the category labels and set category weights of all pixels on the feature images to obtain a set of weighted feature images; encoding the set of the weighted feature images into a feature vector having a fixed length and performing normalization processing the feature vector, and representing a final encoding feature vector of the image using the normalized feature vector; and performing similarity calculation and returning a retrieval result. The method introduces semantic segmentation technology into feature encoding of image retrieval, thereby greatly improving the retrieval effect. The proposed manual design method and the parameter learning method of the deep neural network based on the prior knowledge are effective for obtaining the weight of each category of the image.
Inventors:
LI XIU (CN)
JIN KUN (CN)
JIN KUN (CN)
Application Number:
PCT/CN2018/114826
Publication Date:
December 19, 2019
Filing Date:
November 09, 2018
Export Citation:
Assignee:
GRADUATE SCHOOL SHENZHEN TSINGHUA UNIV (CN)
International Classes:
G06K9/62; G06F16/00
Foreign References:
| CN106909924A | 2017-06-30 | |||
| CN107092870A | 2017-08-25 | |||
| CN106650690A | 2017-05-10 | |||
| US20030133599A1 | 2003-07-17 | |||
| US20100040285A1 | 2010-02-18 |
Attorney, Agent or Firm:
CHINA TRUER IP (CN)
Download PDF:
Previous Patent: HIGHLY ACTIVE STRIGOLACTONE DERIVATIVE, PREPARATION AND USE THEREOF
Next Patent: EARPHONE CAPABLE OF BEING FLEXIBLY WORN
Next Patent: EARPHONE CAPABLE OF BEING FLEXIBLY WORN