Detailed Research

---

Style Separation and Transfer of Korean Portrait Using Principal Component Analysis

Existing studies on style transfer use deep based deep learning methods generated using network models. However, if you can separate the style of the data itself and transfer the style to the content image without using the network, there is a great advantage in terms of time. This paper proposes a method of separating styles from images and transferring styles to content images using principal component analysis.

---

Restoration of JPEG Loss Compressed Image based on Pix2Pix : A Preliminary Study

In the case of the JPEG compression method, there is a process in which each block is independently processed. Since block quantization processing is performed in DCT transformation, a blocking phenomenon occurs when the compression rate is high. Restoring a lost and compressed image to its original image is a necessary technique in modern times. In this paper, we show a preliminary study of restoration using Pix2Pix[2], a GAN-based network. The goal is to restore the converted JPEG image by compressing 95% and 98% based on the PNG image. As a result, the blocking phenomenon of the lost compressed image is removed and the result of improved performance is shown. However, since it can be seen that there is still a difference from the original, it is necessary to improve the performance in the future.

---

Style Interconversion of Korean Portrait and ID Photo Using CycleGAN

There are times when I wonder what a real person would look like if they were painted in a Korean portrait style. Thus this paper proposes a system for Interconversion of Korean portrait and ID Photo using CycleGAN. Experiments that have collected and preprocessed many data sets for high performance show excellent performance converted to style.