An autonomous lightweight model for aerial scene classification under labeled sample scarcity

Article Type

Research Article

Publication Title

Applied Intelligence

Abstract

Aerial scene classification using convolutional neural network (CNN) has gained substantial research interest during last few years. The performance of these deep models is found to improve more when the spatial relationships among the scene features are explicitly modelled using capsule network (CapsNet). However, the combined CNN and CapsNet-based scene classifiers are not only computationally intensive but also often suffer from over-parameterization, leading to remarkable performance deterioration under scarcity of labeled samples. In order to address this issue, we propose a lightweight as well as autonomous CapsNet model which auto-prunes the unnecessary weights/parameters during network learning/training phase, and eventually reduces the computational cost. The efficacy of our lightweight autonomous CapsNet is evaluated after embedding this into PReLim, a recently developed paradigm for remote sensing scene classification under observed sample limitation. Experiments on three benchmark datasets show that our proposed lightweight PReLim (LW-PReLim) is able to attain state-of-the-art accuracy even with 25% less CapsNet parameter count.

First Page

22216

Last Page

22227

DOI

https://10.1007/s10489-023-04694-2

Publication Date

10-1-2023

This document is currently not available here.

Share

COinS