Encoding pairwise Hamming distances of Local Binary Patterns for visual smoke recognition

Abstract To achieve scale invariance, existing methods based on multi-scale local binary patterns (LBP) usually concatenate histograms of LBP codes from different scales. Direct concatenation of histograms is very simple and computationally efficient, but it cannot well model the spatial relationship of LBP codes across scales. Aiming at modeling scale-level variations of LBP codes, we measure and encode the relationship between a pair of LBP codes at the same position from two scales. Gaussian filters are applied to generate the scale space of an image, and original LBP codes are extracted on each scale. The Hamming distance between a pair of LBP codes is used to measure variations of LBP codes across scales. To incorporate the scale-level variations of LBPs into codes, we encode the Hamming distance measures in the same way as LBP to generate a novel code, called Pairwise Comparing Local Binary Patterns (PCLBP). To achieve rotation invariance, LBP and PCLBP codes are aligned to the direction with the maximum magnitude of local differences between a center pixel and its neighbors. To improve reliability of the alignments, we also propose a circular Gaussian smoothing method to remove noise in local differences. Finally, we concatenate the histograms of aligned PCLBP and LBP codes to obtain rotation and scale invariant descriptors for smoke classification. Extensive experiments show that our method obviously outperforms existing LBP variants on smoke datasets, and also achieves outstanding performance on other datasets.