Method of Controlled Image Degradation for Detecting Attention Centers in Classification Neural Networks
Loading...
Date
Journal Title
Journal ISSN
Volume Title
Publisher
ЛНТУ
Abstract
The paper discusses a method of controlled image degradation for detecting attention centers in neural networks for classification tasks. A method is proposed that preserves classification logits during image degradation, using a uniform background and optimizing the ‘transparency’ map. This approach allows working with modern neural network architectures such as MIXER and ViT and improves training by identifying significant elements for classification.
Description
Keywords
computer vision, attention localization, data quality
Citation
Dorenskyi O., Drieiev O., Drieieva H. Method of Controlled Image Degradation for Detecting Attention Centers in Classification Neural Networks // MINTECH 2026. Modern Information Technologies: From Theory to Practice : 1st International Sclentific and Practical Conference : book of abstracts, 22–23 May 2026. Lutsk : LNTU, 2026. P. 883-886.