Enhancement of X-ray battery image quality using deep learning

Abstract

In the burgeoning field of energy storage technologies, efficient and precise quality inspection methods are indispensable for ensuring high-performing, reliable batteries. This study focuses on leveraging advanced deep learning techniques to improve the quality of images used in the battery inspection process. We propose a two-pronged approach that utilizes U-Net, a fully convolutional neural network, and CycleGAN, a generative adversarial network, in conjunction with a diffusion model for generating synthetic data. U-Net, primarily designed for biomedical image segmentation, demonstrates remarkable robustness in enhancing image quality, surpassing the performance of traditional denoising techniques like DnCNN. CycleGAN's inherent advantage of handling unpaired image data effectively addressed the practical challenge of obtaining perfectly matched image pairs. Furthermore, CycleGAN's performance improved significantly when trained with synthetic data generated via the diffusion model, underscoring the benefits of expanding the training dataset with synthetic data generation. Normalization of generated data proved vital in enhancing the performance of CycleGAN, reiterating the importance of appropriate data preprocessing, particularly when merging real and synthetic data. The results from this study present promising implications for industrial applications where high-quality image data is a challenge. The methods used and outcomes achieved could pave the way for more sophisticated deep learning applications in industrial quality inspection processes.