Ancient manuscripts are vital for preserving knowledge of past cultures, social structures, and innovations. However, these documents often suffer from physical damage and text loss over time, hindering their interpretation and the reconstruction of history. This study proposes a deep learning-based model using a dataset of hiragana and kanji for restoring missing characters and evaluates methods to enhance its accuracy. Initially, damaged characters were grouped by type, and restoration performance was analyzed. A decline in accuracy was observed with an increase in character types. To address this, a conditional generative adversarial network was employed, using character labels to better capture target features and improve results. Data augmentation and consistency regularization were also applied to enhance generalization. Combining these techniques improved restoration accuracy, demonstrating their effectiveness in supporting the restoration and preservation of ancient manuscripts.