Randomized Encoding Ensemble: A New Approach for Texture Representation

Although many learning-based approaches have been proposed for texture analysis showing promising results, they use large and complex architectures and suffer from limited data availability for training in real problems. This paper proposes a compact texture representation method based on an ensembl...

Full description

Saved in:
Bibliographic Details
Published inInternational Conference on Systems, Signals, and Image Processing (Online) pp. 1 - 8
Main Authors Fares, Ricardo T., Vicentim, Ana Catarina M., Scabini, Leonardo, Zielinski, Kallil M., Jennane, Rachid, Bruno, Odemir M., Ribas, Lucas C.
Format Conference Proceeding
LanguageEnglish
Published IEEE 09.07.2024
Subjects
Accuracy
Decoding
Ensemble learning
Feature extraction
Image coding
Image recognition
Randomized autoencoder
Texture Representation
Training
Vectors
Online AccessGet full text

Cover

More Information
Summary:Although many learning-based approaches have been proposed for texture analysis showing promising results, they use large and complex architectures and suffer from limited data availability for training in real problems. This paper proposes a compact texture representation method based on an ensemble of Randomized Autoencoders (RAE). In our approach, we process each texture image through multiple RAEs, which perform various random projections in the hidden layer, mapping them into the same dimensional space to learn different image perspectives in the output layer (decoder). We adopt this strategy because the quality of the texture representation can be constrained by a single random projection of the input matrix. Consequently, we propose enhancing feature extraction by concatenating the average of the column values from the learned weight matrices (decoder) in the output layer of each autoencoder. The proposed texture representation was evaluated on four datasets: Outex, USPtex, Brodatz and MBT, showing that our method obtains higher classification accuracies when compared to other literature methods, including deep convolutional neural networks. We also assess the effectiveness of the proposed representation through its application to the practical and challenging task of identifying Brazilian plant species. The results indicate that the proposed texture representation is highly discriminating, showing an important contribution to the texture analysis field and applications.
ISSN:2157-8702
DOI:10.1109/IWSSIP62407.2024.10634030