Dynamic Texture Classification Using Directional Binarized Random Features

Dynamic texture description has been studied extensively due to its wide applications in the field of computer vision. Local binary pattern (LBP) and various variants account for a large part dynamic methods because advantages, such as good discriminability low computational complexity. However, many LBP-based directly extract feature from pixel intensities only use proportion pixels local neighborhood. And their classification performance is usually achieved at cost high dimensionality, which would limit application scenarios. We argue that extracting features gradient domain will capture more discriminative additional directional information, making all neighborhood improve performance. In this paper, we propose simply but effective descriptor inherits advantages LBP while excluding disadvantages. The proposed method consists four stages data processing: 1) gradients extraction; 2) random extraction gradients; 3) hashing features; 4) histogramming. Gaussian first-order derivatives are used filters stable could be generated. Then projection applied each gradients. Both above two conducted via 3D filtering, thus they efficient. Thirdly, binarized encoded into integer codes, histogram built. Finally, histograms concatenated vector. Because 8-bit dimensionality very low. evaluate on three benchmark datasets with test protocols. results demonstrate effectiveness efficiency when comparing state-of-the-art methods.