An Unmanned System for Automatic Classification of Hazardous Wastes in Norway

Separation of waste is an essential step in the recycling process, which can save resources, provide energy and reduce environmental pollution. Separation is a tedious process which is usually done by human workers, who hand-pick the items to separate them. To make this process easier, more accurate and safer, in this work a system is developed that can classify items using image recognition techniques and output classes using a projector. A dataset of 12 (of which eight are combined into a single “others” class, due to them being uncommon) different classes with about 5000 images in total is collected and used to train different classification models using convolutional neural networks and transfer learning. A mean accuracy of 74.043%±12.621% is achieved on test data in 10-fold cross-validation. Unfortunately, the model performs drastically worse on newer data, due to unknown reasons.