Milk powder is the hardest hit area for food adulteration. Among them, infant formula milk powder is expensive and important, with quality being the focus of consumers, manufacturers, and law enforcement agencies. Near infrared-hyperspectral imaging (NIR-HSI) technology combined with chemometrics and machine learning algorithms can detect the content of single adulterant in milk powder. The quantitative prediction of multiple adulterants (melamine, vanillin and starch) in different brands of infant milk powder was studied based on NIR-HSI technology. The hyperspectral images after pixel wise pretreatment were divided into regions of interest (ROI), and the ROI average spectra were extracted. The key variables for modeling were selected based on the classic filtering feature selection algorithms, i.e. Laplacian score (unsupervised) and ReliefF (supervised). Partial least squares (PLS) regression was adopted to establish prediction models. A one-dimensional convolutional neural network (1DCNN) model with a self-defined selection layer was developed. The self-defined layer determined the important wavelength variables according to the multiplicative weight parameters learned after modeling. The root mean square errors of prediction set of Laplacian score-PLS models to predict milk powder, melamine, vanillin and starch were 0.1110%, 0.0570%, 0.0349% and 0.3481%, respectively. The root mean square errors of prediction set of ReliefF-PLS models to predict milk powder, melamine, vanillin and starch were 0.1998%, 0.0540%, 0.0455% and 0.1823%, respectively. The root mean square errors of prediction set of 1DCNN models to predict milk powder, melamine, vanillin and starch were 0.561%, 0.0911%, 0.0644% and 0.2942%, respectively. The first 15 important wavelengths selected by Laplacian score, ReliefF and self-defined selection layer were compared and analyzed. The characteristic wavelength subsets selected by different feature selection methods were different, but the wavelengths near 1210nm, 1474nm, 1524nm and 1680nm were selected in more than one method. The visualization results based on the ReliefF-PLS model demonstrated good predictive ability.