RADAM: Texture recognition through randomized aggregated encoding of deep activation maps

Registro completo de metadados
MetadadosDescriçãoIdioma
Autor(es): dc.contributorUniversidade de São Paulo (USP)-
Autor(es): dc.contributorGhent University-
Autor(es): dc.contributorUniversidade Estadual Paulista (UNESP)-
Autor(es): dc.contributorFederal University of Mato Grosso do Sul-
Autor(es): dc.creatorScabini, Leonardo-
Autor(es): dc.creatorZielinski, Kallil M.-
Autor(es): dc.creatorRibas, Lucas C.-
Autor(es): dc.creatorGonçalves, Wesley N.-
Autor(es): dc.creatorDe Baets, Bernard-
Autor(es): dc.creatorBruno, Odemir M.-
Data de aceite: dc.date.accessioned2025-08-21T15:11:34Z-
Data de disponibilização: dc.date.available2025-08-21T15:11:34Z-
Data de envio: dc.date.issued2025-04-29-
Data de envio: dc.date.issued2023-10-31-
Fonte completa do material: dc.identifierhttp://dx.doi.org/10.1016/j.patcog.2023.109802-
Fonte completa do material: dc.identifierhttps://hdl.handle.net/11449/297583-
Fonte: dc.identifier.urihttp://educapes.capes.gov.br/handle/11449/297583-
Descrição: dc.descriptionTexture analysis is a classical yet challenging task in computer vision for which deep neural networks are actively being applied. Most approaches are based on building feature aggregation modules around a pre-trained backbone and then fine-tuning the new architecture on specific texture recognition tasks. Here we propose a new method named Random encoding of Aggregated Deep Activation Maps (RADAM) which extracts rich texture representations without ever changing the backbone. The technique consists of encoding the output at different depths of a pre-trained deep convolutional network using a Randomized Autoencoder (RAE). The RAE is trained locally to each image using a closed-form solution, and its decoder weights are used to compose a 1-dimensional texture representation that is fed into a linear SVM. This means that no fine-tuning or backpropagation is needed for the backbone. We explore RADAM on several texture benchmarks and achieve state-of-the-art results with different computational budgets. Our results suggest that pre-trained backbones may not require additional fine-tuning for texture recognition if their learned representations are better encoded.-
Descrição: dc.descriptionConselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq)-
Descrição: dc.descriptionCoordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES)-
Descrição: dc.descriptionSão Carlos Institute of Physics University of São Paulo, SP-
Descrição: dc.descriptionKERMIT Department of Data Analysis and Mathematical Modelling Ghent University, Coupure links 653-
Descrição: dc.descriptionInstitute of Biosciences Humanities and Exact Sciences São Paulo State University, SP-
Descrição: dc.descriptionFaculty of Computing Federal University of Mato Grosso do Sul, MS-
Descrição: dc.descriptionInstitute of Biosciences Humanities and Exact Sciences São Paulo State University, SP-
Descrição: dc.descriptionCNPq: #142438/2018-9-
Descrição: dc.descriptionCNPq: #305296/2022-1-
Descrição: dc.descriptionCNPq: #405997/2021-3-
Descrição: dc.descriptionCAPES: #88887.631085/2021-00-
Idioma: dc.languageen-
Relação: dc.relationPattern Recognition-
???dc.source???: dc.sourceScopus-
Palavras-chave: dc.subjectConvolutional networks-
Palavras-chave: dc.subjectFeature extraction-
Palavras-chave: dc.subjectRandomized neural networks-
Palavras-chave: dc.subjectTexture analysis-
Palavras-chave: dc.subjectTransfer learning-
Título: dc.titleRADAM: Texture recognition through randomized aggregated encoding of deep activation maps-
Tipo de arquivo: dc.typelivro digital-
Aparece nas coleções:Repositório Institucional - Unesp

Não existem arquivos associados a este item.