| Resumen |
In the era of big data, we now live on, there is an increasing demand to convert large amounts of scanned documents, such as texts, medical records, and images into digital formats. However, often when scanning introduces imperfections such as salt-and-pepper or background noise, blurring caused by camera motion, watermarking, coffee stains, wrinkles, or faded text. These imperfections carry significant challenges to current algorithms of text recognition, leading to a decline in their performance. To date, a wide range of methods are aimed at reducing noise. This work compares the performance of a CycleGAN model concerning median filter, Wiener filter, adaptive threshold, morphological filtering, and a CNN-based autoencoder. While the CNN-based autoencoder technique gave us the best results, the CycleGAN model approach provided us with comparable results with only 50 training epochs in contrast to the 700 epochs of the CNN-based autoencoder and was superior to the rest of the other contrasted methods. Likewise, data preparation for the training is much simpler in the CycleGAN model due to its property of requiring only unpaired data for training. © 2023 IEEE. |
