Paper Information

According to the U.S. Environmental Protection Agency, only a quarter of waste is recycled, and only 60% of municipalities in the United States have curbside recycling pickup. Plastics have a recycling rate of only 8%, with an additional 16% being incinerated. The other plastics (76%) are disposed of in landfills. Low plastic recycling rates are due to various factors such as contamination and lack of incentive, making it difficult to recycle them effectively. To boost efficiency, automated sorting is crucial. Companies like AMP Robotics and Greyparrot employ optical detection for automated sorting, while MRFs leverage Near IR sensors for precise plastic type detection (1; 2; 3) .

Modern optical sorting, driven by advances in visual detection methods like object recognition and instance segmentation, incorporates modern Machine Learning algorithms. Two-stage detectors like R-CNN (4), Fast R-CNN (5), Faster R-CNN (6), Mask R-CNN (7) and SPP-net (8) employ region proposal and classification leveraging backbone architectures like ResNet (9), CPSNet (10), and EfficientNet (11). Single-stage detectors like YOLO(12), SSD (13), and RetinaNet (14) perform one-pass detection without a separate region proposal stage. While such methods excel under ideal conditions with a large volume of labeled training data, challenges arise in realistic scenarios, emphasizing the need to further examine the efficacy of optic detection for automated sorting.

In this study, we have compiled several novel datasets to evaluate the efficacy of vision-based detection methods. These comprise over 20,000 images sourced from various sources. Through custom and public machine learning pipelines, this study aims to understand how optic recognition works as well as its limitations for sorting. Using metrics like Grad-CAM, saliency maps, and confusion matrices, we highlight the features identified by optic sorting and their limitations. We perform this analysis on our custom trained model from the complied datasets.

To conclude, our findings are that optic recognition methods have limited success in accurate sorting of real-world plastics at MRFs, primarily because they rely on physical properties such as color and shape. Further research is needed to adapt such models to identify plastics more accurately.