Abstract:
In the recent years, there has been an increased trend to digitize the historical manuscripts. This, in addition to preservation of these valuable collections, also allows public access to the digitized versions thus providing opportunities for researchers in pattern classification to develop computerized techniques for various applications. A common pre-processing step in such applications is the restoration of missing strokes and makes the subject of our current study. Historical manuscripts often get degraded due to numerous factors which in turn lead to missing strokes, broken and faded characters. More specifically, we worked on isolated Greek characters extracted from handwriting on papyrus and employ a denoising auto-encoder to reconstruct the missing parts of characters. Incomplete character is given as an input to the denoising autoencoder whereas complete image serves as a target.We conducted a series of experiments to determine the impact of the number of layers and the size of the mask on the efficiency of the denoising auto-encoder in reconstructing the missing regions, particularly if the size of the missing stroke has been potentially increased. To evaluate the results, we used both qualitative and quantitative evaluation metrics, including the root mean squared error (RMSE), structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR). Reported results show an average RMSE 0.04 , PSNR 28.24 and SSIM of 0.90. In the future, we plan to investigate the impact of our proposed preprocessing step on improving other tasks in historical document analysis such as writer identification, character recognition, and writing style classification etc. Specifically, we will explore how the reconstructed missing strokes and characters can enhance the accuracy of these tasks.
