Obrazy s vysokým rozlišením a destilace znalostí v hlubokém metrickém učení s transformátory vidění

Abstract

This thesis investigates the trade-off between performance and complexity with respect to various image resolutions using Vision Transformers in deep metric learning. The objective of metric learning is to use deep neural networks to embed images into representative vectors such that images of the same class cluster together in the feature space while maintaining separation between different classes. Vision Transformers, among various deep architectures, have proven efficient at extracting high-level semantics from diverse image content and are thus employed as the primary models. Nevertheless, high performance is often accompanied by substantial complexity. Knowledge distillation is utilized as an optimization technique to enhance the performance of cost-effective models under the guidance of more complex models. Moreover, image resolution significantly affects the model’s performance. Therefore, this thesis examines the performance/complexity trade-off in asymmetric metric learning, where images are processed at different resolutions. The term resolution refers to either the input image resolution or the resolution of the patches that are separately processed at the processing stage of Vision Transformers.This thesis investigates the trade-off between performance and complexity with respect to various image resolutions using Vision Transformers in deep metric learning. The objective of metric learning is to use deep neural networks to embed images into representative vectors such that images of the same class cluster together in the feature space while maintaining separation between different classes. Vision Transformers, among various deep architectures, have proven efficient at extracting high-level semantics from diverse image content and are thus employed as the primary models. Nevertheless, high performance is often accompanied by substantial complexity. Knowledge distillation is utilized as an optimization technique to enhance the performance of cost-effective models under the guidance of more complex models. Moreover, image resolution significantly affects the model’s performance. Therefore, this thesis examines the performance/complexity tradeoff in asymmetric metric learning, where images are processed at different resolutions. The term resolution refers to either the input image resolution or the resolution of the patches that are separately processed at the processing stage of Vision Transformers.