A completed parted region local neighborhood energy pattern for texture classification.

Binary pattern family has received unprecedented attention due to its primary role in texture representation. Texture images usually have strong region attributes and show dramatic changes from coarse regions to flats regions. However, most methods build texture descriptors by applying the same strategy to all pixels in the whole image, which ignore the difference between coarse regions and flat regions and fail to precisely represent the corresponding region attribute. To solve this problem, this paper proposes a novel completed parted region local neighborhood energy pattern which adaptively extracted discriminative texture information from coarse regions and flat regions. Firstly, we design an adaptive parted region division strategy for feature extraction. Secondly, to provide more rich and discriminative texture features, this paper further develops two novel operators based on the proposed strategy, parted region local difference energy pattern and parted region local pixel energy pattern. Finally, a completed parted region local neighborhood energy pattern is built by combining local sign pattern, parted region local difference energy pattern, and parted region local pixel energy pattern. Extensive experimental results on three popular texture databases demonstrate that the proposed descriptor outperforms recent state-of-the-art methods for texture classification tasks.