PrecisionGAN: enhanced image-to-image translation for preserving structural integrity in skeletonized images

Character skeletonization is a critical aspect of the accurate digital analysis of ancient manuscripts. Our study introduces PrecisionGAN, a generative adversarial network (GAN) specifically tailored for the skeletonization of ancient manuscript characters, marking an advancement in cultural heritage analysis. PrecisionGAN utilizes a U-Net-based generator with innovative multi-head attention and residual connections, combined with a PatchGAN-based discriminator. It is fine-tuned with a hybrid loss function optimized to enhance accuracy and reduce artifacts, even when the training data is imperfect. Our evaluations show that this GAN not only excels in manuscript character skeletonization, achieving superior accuracy and image quality over existing methods but also demonstrates the potential for a range of image processing applications, such as underwater image restoration. Evaluated using the Gaussian similarity measure, our method outperformed the conventional techniques up to 81% of the cases. Further assessments with the structural similarity index measure and a customized metric focused on endpoints and junctions yielded comparable or superior results. This research highlights the versatility of our approach to preserving and interpreting cultural heritage through cutting-edge digital technologies.