Towards more discriminative features for texture recognition

Abstract

Local binary patterns (LBP) are considered to be one of the most computationally efficient descriptor that can also be combined jointly among different variants to increase accuracy. In this study, we propose a method to obtain more discriminative 2D LBP features by optimizing projections of a joint LBP distribution onto the marginal histograms. To find a more efficient representation of the feature vector, we seek the least redundant marginal histograms of a joint LBP distribution via optimizing several constraints. In this way, we aim to have a more compact yet accurate feature vector in contrast to the methods that flatten the joint distribution. Experiments we perform on five popular texture datasets show that the feature vectors optimized with the proposed method provide higher recognition rates with the same size vectors and comparable results even with lower dimensional vectors. We also compare the proposed algorithm to more recent texture recognition methods based on convolutional neural networks and show that it can still provide comparable results even though the resulting feature vectors are smaller by orders of magnitude.