Using Open Data for Training Deep Learning Models: A Waste Identification Case Study

One of the main challenges of building commercial solutions with Supervised Deep Learning is the acquisition of large custom-labeled datasets. These large datasets usually fit neither commercial industries’ production times nor budgets. The case study presents how to use Open Data with different features, distributions, and incomplete labels for training a tailored Deep Learning multi-label model for identifying waste materials, type of packaging, and product brand. We propose an architecture with a CBAM attention module, and a focal loss, for integrating multiple labels with incomplete data and unknown labels, and a novel training pipeline for exploiting specific target-domain features that allows training with multiple source domains. As a result, the proposed approach reached an average F1-macro-score of 86% trained only with 13% tailored data, which is 15% higher than a traditional approach. In conclusion, using pre-trained models and highly available labeled datasets reduces model development costs. However, it is still required to have target data that allows the model to learn specific target domain features.