JOU,JIA FENG
National Taipei University of Department of Computer Science College of Science Student
Email：lucky48tw@gmail.com
LIU, YUAN CHEN
National Taipei University of Department of Computer Science College of Science Professor
Email：liu@tea.ntue.edu.tw
Abstract
In the past, training neural networks was a complex manual work. A large amount of data had to be collected, sorted, and labeled. And then put the data into the neural work for training. But collecting data in a specific field is difficult and expensive. On the contrary, using unpaired datasets and unsupervised architectures for image-to-image translation not only increase the dataset, but also add variety to the dataset. Several studies have proved this data preprocessing method. For example, they adjusted colors and hue distribution but retained the regular attributes as well. This shows that the actual color of the object is crucial, because it affects the final accuracy of the entire model and the final training result. Therefore, the study about imageto- image translation related technology is in demand. For data preprocessing in deep learning, the approach of increasing dataset can enhance accuracy of training.
The study used the image style transfer technology as the basic framework of neural networks, which combined with the loss function of unsupervised cross-domain transfer and migration technology. The feature of one domain was transferred to the feature of another domain. One can flexibly learn from what has been learned from another domain, and then the features were connected together. Therefore, under the condition of one domain, the potential variables could be strengthened and controlled. Specific samples were generated by this transfer technology. The result proved that the output retained more input RGB color features, and the colors became softer than the one produced only by using the image style transfer technology.
Keywords :Generative Adversarial Network, image-to-image translation, image style transfer