부산대학교 도서관

As people's aesthetic becomes more and more diversified, pixel art is loved by more and more people with its unique artistic form and artistic value. In the creation process of pixel image, every pixel needs to be arranged into a zigzag contour according to the rule, and the contour needs to be continuous and clear. However, the existing algorithms often generate more artifacts, and the contour details are not clear. This paper proposed a new unsupervised algorithm for pixel image generation. By using nested U-Net generator structure, multi-scale feature fusion is used to assist in drawing the contour features of pixel image. In addition, structure combination loss is proposed to further ensure the integrity of linear structures such as contour in pixel images. Experimental results show that the proposed method can generate high-quality pixel-style images, and is superior to the existing methods in preserving image details.