Automatic portrait image pixelization

• Published in: Computers & graphics 2021-04, Vol.95, p.47-59
• Main Authors: Shang, Yunyi, Wong, Hon-Cheng
• Format: Article
• Language: English
• Subjects: Algorithms; Clustering; Color extraction; Data analysis; Image abstraction; Image downscaling; Image enhancement; Object recognition; Pixel art; Pixelization; Pixels
• ISSN: 0097-8493; 1873-7684
• DOI: 10.1016/j.cag.2021.01.008

•An automatic portrait pixelization algorithm that can capture portrait features.•Suitable for pixelating natural portrait images with smooth color transition.•Satisfying the requirements of clear edges and limited colors of pixel images.•Producing a unique style between realism and abstraction.•Friendly for untrained people to work with pixel art. [Display omitted] Pixelization is one of the most distinctive image stylization methods, and it is often used in game production, icon design, and advertising printing. One of the most popular examples is portrait pixel art, which retains both the features and imagination of a character. However, the difficulty of pixelization in production and time consumption discourages many untrained people from working in pixel art. In this paper, we propose an automatic pixelization algorithm for portrait images based on Simple Linear Iterative Clustering (SLIC) and Fuzzy Iterative Self-Organizing Data Analysis (FISODATA) algorithms. Several improvements have been made for portrait images using superpixels reordering, cascade object detection, gaussian bilateral filtering, and feature edge enhancement. Our algorithm is particularly well suited for pixelating natural portrait images which have smooth color transition. The proposed algorithm not only satisfies the requirements of clear edges and limited colors of pixel images, but also automatically determines the width and height of the output image and the size of the palette through the image, avoiding intensive human involvement on the resulting image. We compared the results obtained by our proposed algorithm with those of manual pixel art and existing methods. Experimental results show that our algorithm can produce a unique style between realism and abstraction.