Registration or spatial normalization of diffusion tensor images plays an important role in many areas of human brain white matter research, such as analysis of Fraction Anisotropy (FA) or whiter matter tracts. More difficult than registration of scalar images, spatial normalization of tensor images requires two important parts: one is tensor interpolation, and the other is tensor reorientation. Current tensor reorientation strategy possessed many defects during tensor registration. To overcome the shortcomings, we first presented a multi-channel model with one FA and six log-Euclidean tensors, and then proposed an adaptive chaotic particle swarm optimization to find the global minima of the objective function of the multi-channel model. The results on 42 slices inter-subject registration indicate that our proposed method can produce accurate and optimized parameters of tensor registration with fastest speed relative to Genetic Algorithm and Particle Swarm Optimization.