Local patch vectors for rotation-invariant and noise-tolerant texture description

A new texture descriptor is introduced in this paper. In comparison with the traditional LBP-based methods, the proposed method incorporates more global information in Local Patch Distance Vectors (LPDV) and Local Patch Angle Vectors (LPAV) which are the results of distance measurement between vectorized patches. We show that in traditional LBP-based descriptors, with a few levels of quantization, different texture patterns could have exactly the same codes, while in the proposed method, different patterns would have different vectors with continuous entry values. Following LBP-based methods, we perform circular shifting on the LPDVs for rotation invariance, but we show that the circular shifts in LPDV-LPAV do not produce similar codes as in LBP-based methods. For evaluating the proposed method in texture description, two well-known texture datasets are used for classification. Our experiments show that in the presence of noise, the proposed method outperforms the traditional LBP-based methods and in noise-free condition the performance of the proposed LPDV-LPAV is very close to the results of previous works.