Automated Late Fusion of Low Level Descriptors for Feature Extraction and Texture Classification Using Data Augmentation

Feature extraction is an important task for texture image classification. Many descriptors have been proposed in the literature in order to describe textured images locally as well as globally. Researchers' interpretations differ on the effectiveness of these descriptors depending on the field of application, but no one can deny their complementarity. However, fusing different descriptors is not always easy, notably because of their different types (local vs. global, dense vs. sparse ...) and the heterogeneity of the generated features. In this work, we propose to use genetic programming to generate and fuse two different texture classifiers based respectively on HOG and uniform LBP descriptors. Indeed, the proposed method includes a late fusion and data augmentation process in order to combine the classifier's results while using small set of training data. The suggested method benefits from the different information captured by both descriptors while being robust to rotation changes. The performance of the proposed method has been validated on four challenging datasets including different variations. Results show that the proposed method significantly outperforms other low-level methods as well as GP-based methods intended for texture description and classification.