Vision transformers (ViTs) have performed better than convolutional neural networks (CNNs) in image recognition tasks. However, a ViT lacks the inductive bias of a CNN, thus it must be pre-trained on large datasets to guarantee high performance on any other dataset. Recently, networks that introduced a convolutional structure into the input layer of a ViT have been developed, but networks that introduce a convolutional structure to the middle layer of a ViT have been rarely investigated. Here, we propose a ViT with a dilated convolutional structure in the form of source-target attention (STA). By evaluating the image classification performance on the general image dataset (CIFAR10) and remote sensing dataset (EuroSAT) and attention-acquisition ability of the proposed ViT, we demonstrate that introducing the STA into a late layer of ViT achieves better performance and attention as the conventional ViT pre-trained on a large dataset especially for remote sensing dataset, even after pre-training on a small dataset. These results suggest that ViT with a dilated convolutional STA structure in its late-middle layer can efficiently acquire attentions suitable for remote sensing from small datasets.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.