Melanin index is an indicator of the melanin content in the skin. It is important to have an accurate and stable measurement of the melanin index. We utilize a non-contact measurement device to measure diffuse reflectance spectroscopy which combined with machine learning for human skin melanin index detection. First, a non-contact diffuse reflectance spectroscopy measurement device is built and the data is collected. The data is deformed using competitive adaptive reweighted sampling (CARS) and melanin index definitions respectively to prove the rationality of machine learning for spectral data deformation. Then, the performance of machine learning regression models commonly used in predicting melanin indices is compared, and finally a suitable melanin index regression model is selected. The experimental results show that among the machine learning prediction models that combined with the non-contact skin-based diffuse reflectance spectroscopy, the K-nearest neighbor regression model can accurately obtain the melanin index values, the coefficient of determination R² reaching above 0.995 for data validation, and the minimum mean absolute error is 1.251. After comparing the accuracy of five screening wavelength and the dimensionality reduction data obtained by CARS, it is found that the dimensionality reduction data obtained by CARS not only screens out characteristic absorption peaks of different skin chromophore, but also obtains similar prediction accuracy in the prediction models. The aim of this study is to select a suitable prediction model to improve the accuracy of the melanin detection.