A hybrid algorithm combining improved Powell algorithm with genetic algorithm is proposed for terahertz and visible dual- band image registration from human security check apparatus. Using mutual information as similarity measure to overcome the difficulty of extracting feature points from passive terahertz images. The approximate solutions of seven affine transformation parameters are obtained by genetic algorithm and set as the initial point of improved Powell algorithm, and then local search capabilities of improved Powell algorithm are used to obtain the exact solutions of seven transformation parameters. Single genetic algorithm, single Powell algorithm and the proposed hybrid algorithm are respectively used to registrate two groups of terahertz and visible images. Root-mean-square error is used as index to evaluate the registration results.The experimental results show that compared with single genetic algorithm and Powell algorithm, the proposed algorithm can achieve higher matching accuracy. The study can be further used for terahertz/visible dual- band image fusion, to improve the recognition accuracy of concealed object in clothes and enhance the practicability of terahertz human security check apparatus.