Proxy-Based Rotation Invariant Deep Metric Learning for Remote Sensing Image Retrieval

Convolutional Neural Networks (CNNs) are frequently utilized in content-based remote sensing image retrieval (CBRSIR). However, the features extracted by CNNs are not rotationally invariant, which is problematic for remote sensing (RS) images where objects appear at variable rotation angles. Additio...

Full description

Saved in:
Bibliographic Details
Published in:IEEE journal of selected topics in applied earth observations and remote sensing 2024-01, Vol.17, p.1-14
Main Authors: Cai, Zhoutao, Pan, Yukai, Jin, Wei
Format: Article
Language:English
Subjects:
Artificial neural networks
Data mining
Deep neural network
Distance
Embedding
Euclidean geometry
Feature extraction
feature fusion
Image retrieval
Invariants
Learning
Measurement
Metric space
metrics learning
Multilayers
Neural networks
Remote sensing
remote sensing (RS) image retrieval
Remote sensing image retrieval
Retrieval
Rotation
rotation invariant
Semantics
Training
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Convolutional Neural Networks (CNNs) are frequently utilized in content-based remote sensing image retrieval (CBRSIR). However, the features extracted by CNNs are not rotationally invariant, which is problematic for remote sensing (RS) images where objects appear at variable rotation angles. Additionally, because remote sensing images contain a wealth of content and detail information, CNNs may lead to information loss by superimposing multiple convolutional and pooling layers, affecting the ability of the model to extract features. To address these problems, this paper propose a proxy-based feature fusion network (PBFFN). By designing a proxy-based Euclidean distance contrast loss that combines contrast learning within the framework of metric learning, such that the distance between the source image and its rotated image embedding vector in the metric space is closer than any other image, thus endowing the model with a certain degree of rotation invariant. Meanwhile, the global correlation map is generated by multi-layer fusion, under whose guidance the features of each layer are fused to improve the feature extraction capability of the model and to reduce the loss in the image flow process. Extensive experiments based on two public remote sensing datasets show that the method achieves better performance compared to other methods.
ISSN:1939-1404
2151-1535
DOI:10.1109/JSTARS.2024.3382845