A new multi-level radial difference encoded pattern for image classification and retrieval (Q2211693)

The paper proposes an original image descriptor based on multilevel decomposition based on the stationary wavelet transform of the V-channel of HSV representation of the image as most discriminative feature for the texture. The proposed descriptor encodes the local binary patterns of the difference in the wavelet coefficient available at different levels and radii with respect to each individual pixel. Each level of SWT decomposed result is obtained from the smoothened version of the convergence channel to the previous level. The approximation band acts as a source for the SWT wavelet decomposition process. This method selects coefficients available in a high radius at the initial level and coefficients with a lower radius in the next level of the decomposed image. The performance of the proposed feature descriptor is assessed over seven data sets from different domains. Various sampling variants have been investigated, and it has been experimentally demonstrated that the proposed unified template descriptor, together with 8 sampling points, contains the most details for a given image. Different variants of application of proximity measures have been studied, as the highest degree of precision has been achieved with measure D1-distance. As classifier model Random Forest is used, for which the most effective accuracy is achieved with 80 decision trees.