Visible and infrared images are important ways for power inspection robot to detect the health status of power equipment. Image registration can combine the advantages of two types of images and provide a better basis for subsequent status monitoring. To improve the registration accuracy due to the blur of infrared image, this paper proposes a normalized mutual information algorithm based on saliency gradient. First, based on the visual saliency detection of the infrared image, the edge gradient information of saliency area is enhanced. Second, the saliency gradient information and normalized mutual information are combined as a measurement function of registration. Third, to improve the convergence of the image registration algorithm, a cultural wolf pack algorithm is proposed. This algorithm introduces the hierarchical evolutionary characteristics of cultural algorithm into the wolf pack algorithm to establish the belief space and population space. In the iterative process, the evolution of population space is guided by the knowledge of belief space. Finally, the substation inspection image, standard registration test image set, and standard test functions are selected for comparative experiments. The results show that the proposed algorithm has better performance in registration rate and registration speed.