DTG-Net: a deep transfer-generative network for ultrasonic image restoration of deep-seated defects

Ultrasonic microscopy is essential for characterising deep-seated defects in materials. However, deep penetration often reduces imaging resolution because low-frequency probes, although penetrating more deeply than high-frequency probes, have large focal spot diameters that blur defect edges and distort features, hindering accurate defect quantification. To overcome this limit, this paper proposes a Deep Transfer-Generative Network (DTG-Net) for ultrasonic image restoration. DTG-Net is based on a Generative Adversarial Network (GAN) with a localised PatchGAN discriminator to better capture fine defect details. Specifically, the model is pre-trained using an abundant surface dataset acquired by scanning artificial surface defects. Then, a limited set of deep-seated defect images is utilised to fine-tune the pre-trained generator. During this stage, a partial parameter freezing strategy is implemented to prevent overfitting on scarce target data, yielding a restoration model capable of converting blurred deep-seated defect images into their clear counterparts. The method was applied to blurred ultrasonic images of deep-seated inclusions in steel. The results indicate that the proposed method effectively recovers defect edge details, thereby reducing the defect area measurement error from 20.46% to 9.32%. These results demonstrate that the proposed method can improve the quantitative characterisation accuracy of deep-seated defects in ultrasonic inspection.