A Sample Augmentation Method Based on Three-Dimensional Synthesis and Domain-Adaptive Techniques

Transmission line quality defect detection is one of the important application scenarios of artificial intelligence in the power grid field. However, in some cases, the insufficient samples and imbalance problems severely limit the effective training of algorithms, greatly restricting the application scopes of intelligent recognition technologies. In this paper, we designed and implemented a sample enhancement method based on 3D synthesis and domain-adaptive style transfer technology. First, the 3D model of the target to be detected is placed in a 3D scene, and synthetic sample images and annotation information are rendered and output according to preset model and environmental parameters. Then, CycleGAN is used for style transfer on the 3D synthetic sample images, resulting in final sample data for training the power grid quality defect detection model. This method effectively addresses the issue of traditional data augmentation techniques failing to increase sample diversity at the spatial information level through synthetic data generation based on 3D visualization technology. It allows for the flexible, rapid, and efficient generation of any number of annotated synthetic sample data, providing strong support for subsequent model training. On the other hand, the domain-adaptive technique facilitates style transfer, making the 3D synthetic sample data more similar to real data, thereby enhancing the model’s ability in target detection tasks.