Abstract: Aiming at the issues in pointer meter calibration tasks, such as excessive tilt rotation angles and unsatisfactory performance of conventional calibration methods, this paper proposes an improved STN-based image calibration method for pointer instruments. This method employs a front-end network model (ASTN-FP), to predict the homography parameters and pointer angles of meter images. By incorporating an adaptive transformation layer and a feature pyramid structure, it enhances the model’s learning capability for multi-scale meter processing and improves network performance. During the training phase, a Sim2Real training strategy is adopted, where synthetic datasets are used for initial training, followed by fine-tuning with real-world data. In the calibration stage, homography transformation and perspective transformation are combined to strengthen the model’s ability to handle complex transformations. Validation experiments conducted on both simulated and real-world data demonstrate that, compared to mainstream image calibration methods, the proposed method achieves significant improvements in calibration efficiency and average calibration  time, and achieves a recognition accuracy of 95.3% on the calibration  data, verifying the effectiveness of the proposed method.

Key words: instrument calibration, adaptive transformation layer, characteristic pyramid, Sim2Real, synthetic data