Image classification with Fisher Vectors computed with multilevel features extracted from deep filter banks

• Ano de defesa: 2024
• Autor(a) principal: Lyra, Lucas de Oliveira
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Classificação de imagens médicas; Convolutional Neural Networks; Fisher Vectors; Medical image classification; Reconhecimento de textura; Redes Neurais Convolucionais; Texture recognition; Transformadores visuais; Vetores de Fisher; Visual transformers
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45132/tde-18062025-191437/

In this thesis, we investigate the use of Fisher Vector (FV) for encoding multilevel features extracted from deep neural networks in image classification. Specifically, we employ Convolutional Neural Networks (CNNs) and hybrid CNN + Visual Transformer (ViT) models for feature extraction. While CNNs are effective at extracting generalist features, they exhibit a locality bias, which we address using hybrid architectures. The FV encoding method tackles issues related to order-sensitive encoders, such as Fully-Connected layers, in visual texture recognition and related fields (e.g., medical image classification). Our results demonstrate that the proposed approach significantly improves CNN accuracy in visual texture recognition. Despite its usefulness in the context of limited data availability, scalability to larger datasets remains a challenge. To mitigate this, we propose a method for reducing the computational costs of FV encoding. We rigorously evaluate the robustness of this method and apply it to larger datasets within the context of medical image classification. Additionally, we explore the impact of fine-tuning on the models performance. Finally, our approach proves suitable for both small and large datasets, exhibiting competitiveness compared to existing literature.