ACCELERATED NOVEL LOCAL TEXTURE FEATURE EXTRACTION ON A GPU

Texture features benefit object detection, object recognition, content-based image retrieval, and other tasks. Recently, two new local texture descriptors, Threshold Local Binary AND Pattern and Local Adjacent Neighborhood Average Difference Pattern, have been proposed. Graphical Processing Units (GPUs) are instrumental in speeding up many computationally intensive tasks. We have accelerated these texture feature extractors on a graphical processing unit by proposing parallel implementations of the algorithm in this work. Compute Unified Device Architecture (CUDA) has been used to implement the parallel GPU algorithms. We have also optimized the parallelization by leveraging memory hierarchy in a GPU. The results show that we can use GPUs to achieve a speedup of more than 20.