Classification of Biomedical Journal Images using Retargeting-Based Data Augmentation and Visually Explainable Attention Priors
Vinit Veerendraveer Singh (University Of Delaware),* Chandra Kambhamettu (University of Delaware)
The 33rd British Machine Vision Conference
Abstract
Identifying regions of interest in images is vital for solving various computer vision tasks. Convolutional Neural Networks (CNNs) implicitly detect these regions. Per contra, CNN-compatible retargeting-based data augmentation approaches explicitly detect task-critical regions and enhance their spatial coverage. However, these retargeting approaches require modifying the original network architecture and have high space and time complexity. In addition, the task-critical regions learned by these methods can be inaccurate. Techniques that produce visual explanations for decisions from CNNs can faithfully identify task-critical regions, yet, they are primarily used for interpretability purposes. This paper proposes a data augmentation approach that utilizes outputs from visual explanation techniques as attention priors to retargeting-based data augmentations. We evaluated our approach to categorize biomedical journal images in three ImageCLEF datasets. The proposed approach outperformed state-of-the-art data augmentation approaches on these datasets. In addition, our approach has a significantly lower space complexity compared to other retargeting-based data augmentations approaches.Video
Citation
@inproceedings{Singh_2022_BMVC,
author = {Vinit Veerendraveer Singh and Chandra Kambhamettu},
title = {Classification of Biomedical Journal Images using Retargeting-Based Data Augmentation and Visually Explainable Attention Priors},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0497.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by