This paper addresses the need for efficient waste sort-ing strategies in Materials Recovery Facilities to minimise the environmental impact of rising waste. We propose the use of resource-constrained semantic segmentation models for segmenting recyclable waste in industrial settings. Our goal is to develop models that fit within a 10MB memory constraint, suitable for edge applications with limited pro- cessing capacity. We perform the experiments on three networks: ICNet, BiSeNet (Xception39 backbone) and ENet. Given the aforementioned limitation, we implement quanti- sation and pruning techniques on the broader nets, achiev- ing predominantly positive results while marginally impacting the Mean IoU metric. Furthermore, we conduct experiments involving diverse loss functions in order to address the implicit class imbalance of the task: the outcomes indicate improvements over the commonly used Cross-entropy loss function.
A more detailed explaination of our work and our final results can be found on my Portfolio website project page: Semantic Segmentation for Waste sorting