Image registration using modified Local Ternary Pattern

This paper presents an efficient local feature descriptor to register images based on spatial relationship among the neighbourhood pixels. Most traditional feature descriptors in the literature are based on gradient information. Such descriptors are expensive to build. Our proposed descriptor decreases the computational complexity by using a texture operator named Local Ternary Pattern (LTP) and also avoids the gradient calculation. The original LTP method is modified to reduce its dimension so that the proposed local feature descriptor takes less time in matching step. Our experimental results show that the proposed descriptor is significantly more efficient to build than the state of art descriptors such as SIFT, while achieving comparable or better accuracy in registering images against changes in illumination, rotation, view-point angle, image blur and JPEG compression. Furthermore, as our proposed descriptor and SIFT have the same dimensional size, the matching time of the two descriptors is the same.