RETRACTED ARTICLE: Improved transfer learning of CNN through fine-tuning and classifier ensemble for scene classification

In high-resolution remote sensing imageries, the scene classification is one of the challenging problems due to the similarity of image structure and available datasets are all small. Performing training with small datasets on new convolutional neural network (CNN) is inclined to overfitting, and at...

Full description

Saved in:
Bibliographic Details
Published in:Soft computing (Berlin, Germany) Germany), 2022-06, Vol.26 (12), p.5617-5636
Main Authors: Thirumaladevi, S., Veera Swamy, K., Sailaja, M.
Format: Article
Language:English
Subjects:
Accuracy
Artificial Intelligence
Artificial neural networks
Classification
Computational Intelligence
Control
Data Analytics and Machine Learning
Datasets
Deep learning
Engineering
Image resolution
Machine learning
Mathematical Logic and Foundations
Mechatronics
Neural networks
Remote sensing
Robotics
Semantics
Support vector machines
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:In high-resolution remote sensing imageries, the scene classification is one of the challenging problems due to the similarity of image structure and available datasets are all small. Performing training with small datasets on new convolutional neural network (CNN) is inclined to overfitting, and attainability is poor. To overcome this, we go for a stream of transfer learning, fine-tuning strategy. Here, we consider AlexNet, VGG 19, and VGG 16 pre-trained CNNs. First, design a network by replacing the classifier stage layers with revised ones through transfer learning. Second, apply fine-tuning from right to left and perform retraining on the classifier stage and part of the feature extraction stage (last convolutional block). Third, form a classifier ensemble by using the majority voting learner strategy to explore better classification results. The datasets called UCM and SIRI-WHU were used and compared with the state-of-the-art methods. Finally, to check the usefulness of our proposed methods, form sub-datasets from AID and WHU-RS19 datasets with likely labeled class names. To assess the performance of the proposed classifiers compute overall accuracy using confusion matrix and F1-score. The results of the proposed methods improve the accuracy from 93.57 to 99.04% for UCM and 91.34 to 99.16% for SIRI-WHU.
ISSN:1432-7643
1433-7479
DOI:10.1007/s00500-022-07145-1