Title

Fixing Localization Errors to Improve Image Classification

Document Type

Conference Proceeding

Source of Publication

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

Publication Date

1-1-2020

Abstract

© 2020, Springer Nature Switzerland AG. Deep neural networks are generally considered black-box models that offer less interpretability for their decision process. To address this limitation, Class Activation Map (CAM) provides an attractive solution that visualizes class-specific discriminative regions in an input image. The remarkable ability of CAMs to locate class discriminating regions has been exploited in weakly-supervised segmentation and localization tasks. In this work, we explore a new direction towards the possible use of CAM in deep network learning process. We note that such visualizations lend insights into the workings of deep CNNs and could be leveraged to introduce additional constraints during the learning stage. Specifically, the CAMs for negative classes (negative CAMs) often have false activations even though those classes are absent from an image. Thereby, we propose a loss function that seeks to minimize peaks within the negative CAMs, called ‘Homogeneous Negative CAM’ loss. This way, in an effort to fix localization errors, our loss provides an extra supervisory signal that helps the model to better discriminate between similar classes. Our designed loss function is easy to implement and can be readily integrated into existing DNNs. We evaluate it on a number of classification tasks including large-scale recognition, multi-label classification and fine-grained recognition. Our loss provides better performance compared to other loss functions across the studied tasks. Additionally, we show that the proposed loss function provides higher robustness against adversarial attacks and noisy labels.

ISBN

9783030585945

ISSN

0302-9743

Publisher

Springer International Publishing

Volume

12370 LNCS

First Page

271

Last Page

287

Disciplines

Computer Sciences

Indexed in Scopus

yes

Open Access

no

Share

COinS