| Abstrakt: |
Because of its geography and long history, Morocco has seen many populations cohabit or succeed each other on its soil. Consequently, Morocco inherits very rich decorative arts which are traditions and know-how still very much alive. Archaeological sites, mosques, madrasas and palaces, medinas, Berber casbahs, ceramics and multicolored carpets, etc. ... Morocco is rich in a unique artistic tradition. However, given the variety and richness of Moroccan decorative arts, the classification of decorative motifs is a major issue today. Indeed, there are many types of decorations that consist of decorative units borrowed from nature with an abstraction or simplification and some additions or modifications. In this con-text, the classification of motifs according to type is therefore an important aspect for better understanding this art. The objective of this work is to use computer vision approaches for the classification of Moroccan decorative motifs. Thus, three computer vision approaches are compared and evaluated in terms of their performance. For this purpose, two datasets are used with a different level of challenge, one containing texture images and the other containing patterns in complex scenes. The three approaches studied are hand-created features, training a convolutional neural network from scratch and transfer learning. The best results are obtained with transfer learning on both datasets, reaching 95% accuracy. Transfer learning is also generalizable, giving good results in both datasets used. [ABSTRACT FROM AUTHOR] |
