Loading...

Deep Learning Algorithms and Generative Models for Grayscale Image Colorization

Ashouri, Ali | 2022

504 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55147 (02)
  4. University: Sharif University of Technolog
  5. Department: Mathematical Sciences
  6. Advisor(s): Mahdavi Amiri, Nezamoddin
  7. Abstract:
  8. Over the past years, translating a grayscale image to a colorful image, due to its applications in medical imaging, and restoring and colorizing old images, has been popular. The problem we are studying here is automatic grayscale image translation to a colorized image such that the colorized image appears as real as possible. Due to the large degrees of freedom in the allocation of colors to different sections of a grayscale image, this problem is extremely ill-posed. Hence, based on previous works we attempt to utilize Generative Adversarial Networks and Convolutional Neural Networks to overcome this issue. The trained model receives a grayscale image and predicts two chromatic channels of a complete image in Lab color space. Here, we implement three models: (1) Conditional Generative Adversarial Networks with two generators, (2) Regular Conditional Generative Adversarial Networks with one generator, and (3) Deep Convolutional Neural Network. We train these three models on two different data sets and compare the results. The results demonstrate that the convolutional neural network model colorizes grayscale images better than generative adversarial models and also the generative adversarial network with two generators colorizes grayscale images slightly better than the regular generative adversarial network with one generator.
  9. Keywords:
  10. Generative Adversarial Networks ; Deep Learning ; Image Inpainting ; Convolutional Neural Network ; Colouring ; Gray Scale Images

 Digital Object List

 Bookmark

...see more