Anisotropic LBP descriptors for robust smoke detection

Image based smoke detection is a difficult problem especially in open areas since it is heavily affected from its environmental objects. This comes from the transparent property of smoke itself. Therefore, to realize robust smoke detection in such situation, it needs to take into account the effect of the degree of transparency, the change of background objects and so forth. To describe smoke information by image features, they are affected from degree of transparency of smoke, background objects, and other environmental conditions such as the direction and the speed of wind. To address such problems, we apply a novel image feature named anisotropic LBP descriptors, which is considered as a extended variants of LBP. The anisotropic LBP descriptors are simply extended from LBP, which are defined as texture operator using anisotropic neighborhood pixel values. Therefore, they can describe anisotropic deformed image information, which are caused from environmental conditions. To obtain more accurate detection results, we also adopt AdaBoost which uses anisotropic LBP descriptors as input vectors. In this study, each AdaBoost classifier is trained for every anisotropic LBP descriptor and the detection result is obtained from the combined result of those AdaBoost classifiers. We evaluate our presented method and confirm that the our approach works well to obtain robust results.