| Although deep learning approaches have achieved promising results in texture representation, they often require large datasets, involve computationally expensive training procedures, and may suffer from optimization issues. To address these limitations while still leveraging deeper architectures, this paper proposes a texture representation based on a deep randomized neural network trained without backpropagation. The representation is constructed by concatenating summarized learned weights, computed through closed-form solutions, from each hidden layer of the network. This strategy enables the model to capture complementary features across layers, enhancing the descriptive power of the resulting representation. In addition, a multi-scale analysis is incorporated by extracting image patches of different sizes, allowing the model to capture both micro- and macro-textural patterns. The proposed approach was assessed on four datasets (Outex, USPtex, Brodatz, and MBT) achieving competitive results compared with several existing methods, including deep convolutional neural networks. The results indicate that the proposed method provides an effective and computationally efficient solution for texture representation in computer vision and pattern recognition tasks. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.