Handwriting ink is an important physical evidence of the identification in judicial appraisal. In order to improve the accuracy of ink inspection, we employed Raman spectroscopy for the non-destructive inspection of ink samples. First, the pre-processed spectral data were dimensionally reduced to construct a model of partial least squares discrimination analysis. Then, after the prediction effect was verified by the area under the receiver operating characteristic curve, 36 feature variables with the highest variable importance for the projection were extracted. Furthermore, the feature variables were input as data into a multi-layer perceptron with 13 neurons in the hidden layer, and the final training accuracy rate was 87%, without overfitting. We also found that combining the feature extraction of variable importance for the projection with supervised multi-layer perceptron training could effectively compress the data and shorten the analysis time. Besides, the connection weight between perceptron layers could be adjusted through autonomous learning, which improved the credibility and accuracy of handwriting ink classification results.