SYMMETRY CLASSIFICATION OF REGULAR TEXTURES USING CONVOLUTIONAL NEURAL NETWORKS AND MACHINE LEARNING TECHNIQUES

Abstract

In the context of texture analysis and classification, the classification of regular textures into 17 distinctive wallpaper patterns represents a significant advance. This field is particularly relevant in areas such as art and decoration. This research uses advanced computer vision techniques to automate the classification process, using a compiled dataset of 1,700 images divided into the 17 classes. To address the challenge, two innovative approaches designed to exploit the texture representation of convolutional neural networks (CNNs) are proposed. These approaches are based on transfer learning. The first approach features a shallow CNN with four convolutional layers with the aim to capture texture patterns from middle layers. The second strategy is a hybrid approach that combines traditional texture descriptors (GLCM, LBP, and Gabor filters) with deep features extracted from the ninth layer of the VGG19 model. These methodologies have demonstrated considerable success, achieving accuracies of 78% and 80%, respectively.