Analysing which Machine Learning Algorithms Segregate Industrial Plastic Waste Most Efficently—a Comparative Study of SVM, Logistic Regression and Decision Tree Classifiers on ResNet-18 Embeddings.

The accumulation of plastic waste worldwide poses critical environmental and economic challenges. An Accuracy-Speed Gap in industrial material recovery facilities exacerbates them. Manual sorting is hazardous and unsustainable, as traditional mechanical sensors often misclassify visually identical polymers. This study evaluates a custom-made hybrid deep learning framework that compares the classification accuracy of three machine learning algorithms applied to deep visual embeddings for the segregation of industrial plastic waste. The methodology differentiates visual feature extraction from mathematical classification. This is by utilising a ResNet18 Convolutional Neural Network to process RGB image data into 512-dimensional continuous numerical arrays. These high-dimensional embeddings were subsequently fed into 3 optimised classical machine learning algorithms: SVMs, LR and DTC. The models used were trained using a dataset of 7 distinct chemical polymers. The empirical results revealed that the SVM outperformed both LR and DTC. Thereby, achieving an aggregate accuracy of 82.9% and high precision in identifying LDPE. However, the Logistic Regression and DTC models scored much lower than the SVM models. Thus, it reflects the limitations of linear and orthogonal logic in separating dense spatial arrays. This research has demonstrated that replacing computationally intensive deep classification layers with regularised SVM hyperplanes can advance the industrial viability of chemical segregation.