Scene graph generation method based on super relation learning network

A technology for learning networks and super-relationships, applied in neural learning methods, biological neural network models, instruments, etc., it can solve the problems that no method considers high-level connections, and the performance impact is particularly significant, and achieves efficient distributed training. Significant, high inference and integration effects

Active Publication Date: 2021-07-02
HANGZHOU DIANZI UNIV +1
View PDF9 Cites 3 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0010] (3) Currently, there is no method that considers high-level connections of relations, i.e. transitive inference
Howev

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Scene graph generation method based on super relation learning network
  • Scene graph generation method based on super relation learning network
  • Scene graph generation method based on super relation learning network

Examples

Experimental program
Comparison scheme
Effect test

Embodiment Construction

[0098] The detailed parameters of the present invention will be further specifically described below.

[0099] Such as figure 1 , 2 , 3 and 4, the present invention provides a scene graph generation method (HLN) based on Hyper-relationship Learning Network (Hyper-relationship Learning Network).

[0100] The target frame generation network described in step (1) is as follows:

[0101] 1-1. The backbone network adopts the method of ResNetXt-101-FPN. The feature dimension extracted by the backbone network is 256. And the effect of data enhancement is not used, that is, the way of image flipping is not used.

[0102] 1-2. The candidate frame size of the Region Proposal Network is selected from (32,64,128,256,512). Each layer of FPN in training and testing selects 1000 candidate boxes, a total of 4 layers. The feature dimension of the frame extraction after the region generation network is 256.

[0103] 1-3. During the sampling process of the target frame, 64 target frames a...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

No PUM Login to view more

Abstract

The invention discloses a scene graph generation method based on super relation learning. The method comprises the following steps: 1, target interaction is enhanced through a target self-attention network, and features of targets are fused; 2, interaction between the target and the relation is enhanced through the target-relation attention network, and features between the target and the relation are fused; and 3, transmission reasoning of the super relation is integrated through the super relation attention network. and 4, model training is performed, a target loss function and a relation loss function are put into an optimizer, and gradient return and updating are performed on network parameters through a back propagation algorithm. The invention provides a deep neural network for scene graph generation, and particularly provides a super relation learning network, interaction and transmission reasoning between a target and a relation are fully utilized, the reasoning capability of the relation in scene graph generation is improved, and the performance in the field of scene graph generation is greatly improved.

Description

technical field [0001] The present invention proposes a scene graph generation method (HLN) based on Hyper-relationship Learning Network. HLN originates from a hypergraph, using a multi-layer attention network to interact between objects. We further propose the Object-Relation Attention Network (OR-GAT) to autonomously fuse features from the interactions between objects and relations. The present invention proposes a super-relational attention network for the first time to integrate super-relational transfer reasoning. where super-relations refer to a subset of the relations among the three objects. Through the interaction between objects, the interaction between objects and relations, and the transitive reasoning of super-relations, the predictive effect of relations is obviously improved. Background technique [0002] Scene Graph Generation (SGG) aims to detect objects and predict object relationships. These detected objects and relationships then constitute the scene ...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Application Information

Patent Timeline
no application Login to view more
IPC IPC(8): G06K9/62G06N3/04G06N3/08
CPCG06N3/084G06N3/047G06N3/044G06N3/045G06F18/24
Inventor 俞俊陈志刘晓鹏张健张驰詹忆冰
Owner HANGZHOU DIANZI UNIV
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap