Combining Local and Global Descriptors Through Rotation Invariant Texture Analysis for Ulos Classification

Abstract

Performing images augmentation for data multiplication, to some degree, has negative impact to classification tasks, particularly to the object that has texture patterns with specific direction (anisotropic). As Ulos data mostly are anisotropic textures, the convolution neural networks (CNNs) fail to discriminate image if the images are arbitrarily rotated. This is due to CNNs are not rotation invariant. To benefit anisotropic and isotropic (has no specific direction) textures, conducting features extraction with discrete techniques is needed. Extracting features by wavelet transform (DWT) for directional specific patterns changes the wavelet energy features significantly while the isotropic one does not. To address the issue radon transform is first being employed, as to get principal direction for anisotropic textures. The output of wavelet transform is just globally rotation invariant. On this work, we propose a new approach to obtain robust features set by combining both local and global rotation invariant, as the output from LBP-ROR and wavelet transform. Our work shows that the performance outperforms the previous research done by scholars.