عمر هاتف محمد عبدالرحمن
  • UNIVERSAL IMAGE AND AUDIO RESTORATION USING DEEP LEARNING
  • Restoring the original image or audio signal from a distorted version is a challenge in a real-world application; Traditional techniques such as the Wiener filter and statistical approach have been used, but recently deep learning has been widely found in many applications due to its high-performance quality. The main objective of this paper is to present a new algorithm to restore the images and sound signals suffering from different types of distortion, including blurring, noise, and other degradation processes. The degradation process is identified using the convolutional neural network VGG16, while the conditional-GAN is used for restoration due to the type of the identified distortion. The algorithm also successfully restored audio signals by converting the 1D audio signal to 2D image-like, using short-term fast Fourier transform (STFT). For training and evaluating both distortion identifier and restoration process, 31080 images and 1132 speech signals are tried.