Deep learning framework for image enhancement of phased array ultrasonic imaging

• Proposed model translates S-scan images to high-resolution TFM images. • Proposed model combines high-quality imaging with rapid processing. • Coordinate information reflects spatial relationships between transducer and ROI. • Computational time is reduced by 3-fold compared to conventional TFM methods. • Model maintains performance when fine-tuned with minimal experimental data. Phased array ultrasonic imaging is widely used in non-destructive testing for defect detection. Sector scan (S-scan) is widely used method for rapid inspection at lower resolution. In contrast, the total focusing method (TFM) offers high-resolution images, making it effective for the accurate characterization of defects. This study proposes a deep learning framework for rapid high-resolution imaging by transforming S-scan into TFM-quality images. The proposed neural network generates enhanced visualization from S-scan data by integrating the spatial coordinate information of each image patch relative to the phased array transducer. On simulated images of crack-like defects, the results demonstrate low mean absolute error and high structural similarity, indicating that it achieves high fidelity with the ground-truth TFM image. In addition, the image reconstruction is approximately 3 times faster compared to conventional TFM, highlighting the potential of the proposed method for rapid inspections at higher resolution. Moreover, an aluminum block specimen with artificial defects was fabricated to evaluate the robustness of the proposed model, confirming that the performance is maintained when the pre-trained model is fine-tuned with a small amount of experimental data. Therefore, this framework presents an effective method for accurate and cost-effective ultrasonic inspection by combining the rapid scanning capability of S-scan with the high-resolution of TFM.