Enhancing Semiconductor X-RAY Images: A Framework Combining Denoising and Super-Resolution Modules With a Novel Dataset

Semiconductors are essential components of modern technology, and inspecting their defects has become increasingly important. One popular approach is to inspect them using X-ray images, but the release of semiconductor images to the public is restricted due to security concerns. This limited dataset poses a challenge, particularly when developing learning-based inspection methods. Another problem with X-ray images is the noise and low resolution, which lowers the inspection accuracy of machine and human inspectors. To overcome this challenge, we introduce a new public semiconductor dataset and propose a new X-ray image enhancement framework to help inspectors increase inspection accuracy. Precisely, we have collected a new dataset consisting of three types of X-ray images obtained from actual industrial settings. Additionally, our enhancement method can improve all types of images without the need to train a separate model for each type. This is achieved by analyzing the noise within the dataset and building on previous studies, and using objective functions defined in the frequency domain. Various experiments and ablation studies have shown that our method can significantly improve X-ray semiconductor images, both qualitatively and quantitatively. Our proposed semiconductor dataset is publicly available at this URL: https://drive.google.com/drive/folders/1_YSuxR_YGNah1mnPGNd-pgrRULNZP5gn?usp=sharing.