Eric Julianto, Dian Alhusari, Safrizal Ardana Ardiyansa
Dominant color extraction is a critical task in computer vision, yet standard algorithms like K-Means often fail when background elements occupy the majority of the image area. This paper proposes a robust approach that integrates Spectral Residual Saliency with a novel Power-Law Modulation scheme (p=4) to aggressively prioritize visually important regions. By transforming the pixel space based on saliency weights before clustering, our method ensures the generated palette represents the primary subject rather than background noise. Through extensive experiments on a large-scale dataset of 750 high-resolution images, we demonstrate an average 11.6% improvement in foreground coverage over baseline K-Means, with a significant 19.7% gain in complex advertising layouts. Furthermore, validation via a deployed web-based tool confirms that the proposed method aligns more closely with human perceptual judgments. © 2026 IEEE.
Braincore Indonesia, Indonesia; Universitas Esa Unggul, Department of Computer Science, Jakarta, Indonesia; Universitas Brawijaya, Department of Mathematics, Malang, Indonesia
