Performance evaluation of classical classifiers and deep learning approaches for polymers classification based on hyperspectral images

Plastics are very valuable material for their desirable characteristics being one of them, their durability. But this characteristic turns plastics into an environmental problem when they end in the environment, and they become one source of contamination that can last for centuries. Thus, the first step for effective recycling is to identify correctly the types of plastics. In this paper, different classical classifiers as Random Forest, KNN, or SVM are compared with 1-D CNN and LSTM to classify plastics from hyperspectral images. Also, Partial Least Squares Discriminant Analysis has been included as the baseline because is one of the most widely used classifiers in the field of the Chemometrics community. The images were preprocessed with several techniques as Standard Normal Variate or Savitzky-Golay Polynomial Derivative to compare their effectiveness with raw data with the classifiers. The experiments were carried out using hyperspectral images with a 240 bands spectrum, and six types of polymers were considered (PE, PA, PP, PS, PVC, EPS). The best results were obtained with SVM+RBF and 1-D CNN with an accuracy of 99.41% and 99.31% respectively, preprocessing the images previously with Standard Normal Variate. Also, PCA and t-SNE methods were tested for dimensionality reduction, but they don’t improve the classifier performance.