Texture Classification using Multiple Local Descriptors

Classifying images based on texture features is an active topic in computer vision and pattern recognition field. Many applications like biomedical image analysis, image retrieval, and face recognition emerged from texture classification task. In this paper, we propose a new method to classify texture images by combining multiple histogram-based texture descriptors. First, we compute new efficient features called Joint Motif Labels (JML) and Motif Patterns (MP) descriptors. Both descriptors are based on the motif Peano scan concept that traverses image pixels in a 2×2 neighborhood producing one of 12 motif patterns, according to certain criteria. JML uses additional information, mean and variance, as joint distribution with motif patterns. After that, texture descriptors like Rotation Invariance Co-occurrence Among Local Binary Pattern (RIC-LBP) and Joint Adaptive Median Binary Pattern (JAMBP) are combined along with the new JML and MP descriptors in order to improve the classification performance. Experiments are performed on challenging texture datasets namely, KTH- TIPS-2b and DTD using two classifiers, kNN and SVM. The experiments demonstrate that our approach performs better than the single best texture descriptor with an accuracy of 67.2% and 43.5% on both datasets respectively.