Implementasi Vision Transformer (VIT) dan K-Means Clustering dalam Klasifikasi Gaya Visual Poster untuk Otomatisasi Ekstraksi Skema Warna Digital

Abstract

The purpose of this study is to implement the Vision Transformer (ViT) architecture and K-Means Clustering algorithm to automate fine-grained visual style classification of posters and the extraction of digital color schemes. MSMEs and beginner digital creators often face challenges in determining harmonious color schemes for promotional poster designs, leading to a time-consuming trial-and-error design process. The research utilized a custom dataset of 1,011 poster images scraped from the Behance platform, evenly distributed into three visual style classes: Bauhaus, Swiss Style, and Minimalist. The data preprocessing employed a dual-pipeline approach, where color extraction was optimized through SLIC Superpixel segmentation within the CIELab* perceptual color space to significantly reduce computational complexity without degrading the boundaries of original visual elements. The evaluation results demonstrated that the ViT model performed exceptionally well, achieving an accuracy of 79% and a Macro F1-Score of 0.80. In the color extraction module, the K-Means algorithm successfully formed five highly separated dominant palette clusters with an average Silhouette Score of 0.7698. The perceptual color accuracy validation recorded an average  of 4.7648, proving that the extracted hexadecimal (HEX) palette codes operate well within the acceptable limits of human visual deviation.