How to exploit large image data in the fields of texture classification: A case study with local binary patterns

In the fields of texture classification, the sizes of images significantly vary according to the respective classification scenario. Whereas quite small image patches mostly lead to good classification accuracies, increasing the image size sometimes even has a negative effect. In this work, we focus on derivatives of Local Binary Patterns as these feature extraction methods offer a high discriminative power and efficiency on the one hand an can be effectively analyzed on the other hand. The aim is to get new insight and furthermore to explore strategies which can help to increase the classification performance. We investigate these strategies which exploit the obviously high distinctiveness of small image patches and simultaneously the redundancy available in large image patches. Finally it can be concluded that the traditionally applied strategies for texture classification should be reconsidered in case of sufficiently large image data.