Medical, Scene And Event Image Category Recognition Using Completed Local Ternary Patterns (CLTP)

Many of texture descriptors are proposed based on the Local Binary Pattern (LBP) and have been achieved remarkable texture classification accuracy such as Completed LBP (CLBP) and Completed Local Binary Count (CLBC). However, the LBP suffers from two weaknesses where: 1) it is sensitive to noise and; 2) it sometimes classify two or more different patterns falsely to the same class. To overcome the LBP weaknesses, we propose a new texture descriptor which is defined as Completed Local Ternary Pattern (CLTP). The CLTP was used for rotation invariant texture classification. It demonstrates superior texture classification accuracy as compared to CLBP and CLBC descriptors. This is because, the CLTP is more robust to noise and has a high discriminating property that achieves impressive classification accuracy rates. In this paper, two types of experiments are carried out. In the first experiment, different amount of additive Gaussian noise is added to the TC10 Outex texture data set to investigate and prove the robustness of the CLTP against the noise. For the second experiment, the performance of CLTP for image category recognition is studied and investigated. A variety of image datasets are used in the experiments such as scene data set (e.g., Oliva and Torralba datasets (OT8)), Event sport datasets, 2D HeLa medical images, and our new scene data set, defined as USM scene data set. The experimental results proved the superiority of the CLTP descriptor over the original LBP, and different new texture descriptors such as CLBP in the image category recognition, as well as the robustness against the noise. In 2D HeLa medical images, the proposed CLTP has achieved the highest state of the art classification rate reaching 95.62%