Transformer-Based Super-Resolution Scanning Acoustic Imaging for Industrial Inspection

Despite significant advancements in super-resolution (SR) techniques, their application in industrial nondestructive testing (NDT) remains challenging due to stringent requirements for accuracy and reliability. Among NDT technology, scanning acoustic microscopy (SAM) is widely used in semiconductor industries to generate high-resolution images for detecting defects in increasingly miniaturized and complex components. The integration of SR into SAM has the potential to enhance image quality without increasing scanning time, making it a promising solution for industrial applications. However, a main problem in achieving optimal SR performance in this field is the limited availability of training data, as scanning images of the electronic samples are often confidential. We introduce a method for generating synthetic low-resolution datasets from high-resolution images using a high-order degradation model to mitigate this data scarcity problem. To address the demand for fast, high-resolution imaging in SAM, we present an SR approach that combines Generative Adversarial Networks (GANs) with Transformer-based architectures, leveraging Transformers as the generator network to enhance the GAN model's performance. Our model, named SAMTransGAN, incorporates transfer learning and a dynamic loss function that balances pixel-wise, perceptual, and adversarial loss for optimal results. Our SAMTransGAN achieves better performance, attaining Peak Signal-to-Noise Ratio (PSNR) of 28.158 dB and a Structural Similarity Index Measure (SSIM) of 0.835 on the testing dataset, surpassing all other compared models. We also present a possible deployment framework for our proposed SAMTransGAN on the web page as an efficient and compact SR solution for the NDT industry. Our model in-creases image resolution while preserving critical feature details, demonstrating its practical applicability in industrial settings.