An optimization method for two-dimensional LBP feature vectors

Local binary patterns (LBP) is considered to be one of the most discriminative and computationally efficient descriptor for many computer vision applications. Among numerous variants of LBP, there are also approaches that construct 2-dimensional (2D) histograms to provide a better representation of texture patterns. Those approaches obtain final feature vector by either concatenating marginal histograms of 2D distribution; or flattening the whole distribution in a higher dimensional vector. The resulted feature vector is a more compact one in the former scenario, however, the vector in the latter can provide better accuracy. In this thesis, we propose a method to make LBP features more discriminative by optimizing projections of joint LBP distribution onto the marginal histograms. In order to find a more efficient representation of the feature vector, we seek for the least redundant marginal histograms of a joint LBP distribution via optimizing several constraints. In this way, we aim to have a more compact feature vector in contrast to the methods which flatten the joint distribution without sacrificing accuracy. Experiments we perform on five popular texture datasets show that the proposed method provides higher recognition rates with the same size feature vectors and comparable results even with lower dimensional vectors.