2026 年 63 巻 1 号 p. 59-65
Accurate color reproduction is a critical factor in industries such as printing, packaging, and textile manufacturing. While spectrophotometers provide precise reference measurements, they are expensive, complex, and unsuitable for low-cost, rapid applications. Smartphone-based colorimetry, leveraging machine learning (ML), is emerging as a viable alternative. However, powerful models like XGBoost often suffer from severe overfitting when applied to the small, highly nonlinear datasets typical of color calibration. This study first quantifies this problem using a baseline XGBoost model, which achieved excellent training performance ΔE00 < 0.5 but failed to generalize, yielding severe color differences on test data (mean ΔE00 ∼ 11.0-15.0). To address this limitation, we propose a hybrid architecture combining K-Means clustering and Support Vector Regression (SVR). This approach first utilizes K-Means to partition the color space into localized, similarity-based clusters. Subsequently, specialized SVR models are trained for each L*, a*, and b* component within each distinct cluster. Experiments were conducted on a dataset of 92 PANTONE color patches captured under three lighting conditions (D65, D65+UV, and Indoor Fluorescent). The hybrid K-Means+SVR model demonstrated superior generalization, reducing the average color difference to a mean ΔE00 between 0.80 and 1.02 across all conditions—a range considered virtually imperceptible to the human eye. These findings confirm that integrating K-Means to localize the feature space enables SVR to effectively model nonlinear relationships, successfully mitigating overfitting and providing a robust and accurate color calibration solution for mobile devices.
