Abstract

Climate change and environmental degradation mark the urgent need for sustainable resource management, with recycling playing a central role. To improve efficiency in recycling processes, this thesis investigates whether image recognition and machine learning can be used to classify the contents of recycling bags based on images. The study focuses on two tasks: first, determining whether a bag contains single or a mix of materials, second, identifying the specific material category present. Deep learning models based on the ResNet-18 and VGG16 architectures were trained on real images of recycling bags across 22 categories, including both pure and mixed material classes. Three datasets were used, with variations in size and class balance to evaluate performance under different conditions. The best binary classification model, a VGG16 network, achieved a test accuracy of 83.47%. It was relatively reliable at identifying bags with a single material but less so for bags with mixed content. This result indicates that the model works well for simpler cases but struggle with more complex ones. A more balanced dataset improved VGG16 performance, while ResNet-18 was less sensitive to class imbalance. In addition, dropout regularization proved more effective than batch normalization in this context. For multi-class classification, some categories were predicted accurately while others were misclassified entirely. The results suggest that image recognition has potential to support automated waste classification. However, further development is needed. Particularly more training data, deeper networks and robustness to varying conditions are necessary for model improvement. With these changes, potentially improved models could reduce manual labor, increase classification accuracy and improve efficiency of recycling management.