A weakly supervised approach for recycling code recognition

Waste sorting at the household level is a virtuous process that can greatly increase material recycling and boost the circular economy. To this purpose, waste must be differentiated by material (e.g., PVC, Polyethylene, Paper, Glass, Aluminum, etc.), a task that can be simplified by printing a recycling code on the product case. Unfortunately, the large number of recycling codes printed on products makes this process unfriendly for many users. In this work, we propose a vision-based mobile application to support users in recognizing recycling codes for proper waste sorting. The proposed system combines a dual-head CNN with an image processing pipeline (based on domain knowledge) in order to improve: (i) the reliability of symbol detection/classification and (ii) the weakly-supervised labeling of new samples during iterative training. Our experimental results prove the feasibility of developing effective applications with minimum effort in terms of data collection and labeling, which is one of the main obstacles to successfully applying deep-learning techniques to real-world problems.