DENOISING CITRA TULISAN TANGAN AKSARA LAMPUNG MENGGUNAKAN CONVOLUTIONAL AUTOENCODER

Abstract

The history of a nation or a region is stored in written historical documents using paper, walls, stone, metal, and other media. Efforts to maintain cultural heritage, including these documents, are still being carried out. One of the efforts is to save it in digital form or photos, but it is possible for the obtained images become noisy images. Many factors caused an image to have noise including outdated documents, image results that are influenced by camera lenses, lighting that is not ideal, ect. Noise can affect the information in the image, it is needed to made improvements so that the quality image results can be used for other purposes, both as digital documents and further research such as written recognition. In this research, the Convolutional Autoencoder approach is used to study noise from training data and reconstruct the image into a noise-free image. The noise used in this study will be created using the Gaussian, Salt & Pepper, and Spackle methods on the image of the Lampung script. The hyperparameters on the Convolution Encoder that were tested produced good performance for the model used by achieving low loss of 0.1453 and vall_loss of 0.1504 and also could reduce noise contained in images with various noise types and intensities.