Synthetic aperture radar (SAR) has become an important data source in the field of object recognition owing to its high resolution and all-weather characteristics. The traditional data expansion method has difficulty increasing the diversity of samples, which limits the promotion and application of SAR data. Therefore, in view of the shortcomings of traditional SAR data augmentation methods, such as insufficient diversity and poor practicability, we proposed a new idea that can generate samples from different angles. First, Lee filtering and edge direction gradient algorithms are combined to construct a multiscale recursive template matching model, which can identify the target azimuth accurately. Second, we constructed an Angle-Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets (Angle-InfoGAN) model for data generation and extended the original datasets with different new angles. Finally, we applied this method successfully to Moving and Stationary Target Acquisition and Recognition (MSTAR) datasets, and the Fréchet inception distance (FID) was used to compare other data enhancement models to validate the performance of the Angle-InfoGAN model. The samples generated by the Angle-InfoGAN model effectively improve the scale and diversity of SAR image datasets and lay a solid data foundation for deep-learning-based SAR object detection.