Comparative Analysis of Transfer and Continual Learning for Vision-Based Particle Classification in Plastics Sorting for Recycling

We use the publicly available computer vision research dataset CIFAR100 to assess model performance, calibrate the model hyperparameters, simulate and explore transfer and continual learning techniques. According to our results, the main factors affecting test accuracy are the following: calibration of hyperparameters, such as the learning rate (up to +13%), bigger batch sizes of only new data (up to +9%) or mixed with previous data with more frequent retraining phases (+4%) and differing continual learning strategies (±3%). Based on these assessments, we provide evaluations on our DWRL dataset, which features plastic trash particles on a moving conveyor belt and provide recommendations for the implementation of continual learning frameworks in the context of plastics recycling.