Abstract: In order to realize the complementary advantages of different optical modalities and to help power equipment fault detection and location tasks, this paper uses visible images to enhance the texture information of infrared images. The existing infrared-visible image registration technology is difficult to accurately align local fine structures of power equipment, thus, to address this issue, an adaptive registration algorithm with supervision and retraining (ARSR) is proposed for the first time. The algorithm mainly includes a dual order anisotropic Gaussian directional derivative (Dual-AGDD) mechanism and a double-view matching parameter retraining (DVMPR) framework. Firstly, Dual-AGDD is proposed to filter and orient feature points. The 1st-AGDD performs adaptive local refinement corner detection for power equipment, and the 2nd-AGDD constructs a Gaussian feature triangle to determine the main direction of the feature point. The local intensity invariance method is adopted to construct feature descriptors. Then, the DVMPR framework is proposed to correct and constrain the perspective scale and view rotation of images. Finally, based on the 3σ principle, support vector regression (SVR) is improved to eliminate the mismatched points and complete the registration of heterogeneous data. The experimental results show that, when heterogeneous images of power equipment with different rotations and scale differences or in different environments are registered, the average RMSE of the algorithm in this paper is 2.65, and the average precision is 98.57%. The results exhibit strong rotation invariance, scale invariance, and environmental robustness, and are significantly better than those of existing registration algorithms such as CAO-C2F and SuperPoint-SuperGlue algorithms. The algorithm can improve the registration accuracy of heterogeneous images of power equipment with fine structures.