Detailed Evaluation of Modern Machine Learning Approaches for Optic Plastics Sorting

TL;DR

The paper addresses the effectiveness of vision-based optical sorting for plastics in real-world recycling streams. It assembles a large, diverse image dataset and evaluates multiple ML models (e.g., Mask R-CNN, ResNet-34, Xception) with interpretability analyses (Grad-CAM, saliency maps) to understand what visual cues drive decisions. Findings show that, in real-world settings, these models often rely on superficial cues like color, shape, or packaging labels rather than intrinsic material properties, with activations frequently focusing on conveyors rather than plastics. The work argues for multi-modal sorting approaches—combining vision with spectroscopy or embedded identifiers—and provides a valuable, public dataset resource to advance robust plastic sorting research and practice.

Abstract

According to the EPA, only 25% of waste is recycled, and just 60% of U.S. municipalities offer curbside recycling. Plastics fare worse, with a recycling rate of only 8%; an additional 16% is incinerated, while the remaining 76% ends up in landfills. The low plastic recycling rate stems from contamination, poor economic incentives, and technical difficulties, making efficient recycling a challenge. To improve recovery, automated sorting plays a critical role. Companies like AMP Robotics and Greyparrot utilize optical systems for sorting, while Materials Recovery Facilities (MRFs) employ Near-Infrared (NIR) sensors to detect plastic types. Modern optical sorting uses advances in computer vision such as object recognition and instance segmentation, powered by machine learning. Two-stage detectors like Mask R-CNN use region proposals and classification with deep backbones like ResNet. Single-stage detectors like YOLO handle detection in one pass, trading some accuracy for speed. 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 compiled novel datasets totaling 20,000+ images from varied sources. Using both public and custom machine learning pipelines, we assessed the capabilities and limitations of optical recognition for sorting. Grad-CAM, saliency maps, and confusion matrices were employed to interpret model behavior. We perform this analysis on our custom trained models from the compiled 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.

Figures (6)

• Figure 1: The architecture of Mask RCNN he2017MaskRCNN with its components, a popular Deep Learning architecture used to identify and classify multiple objects of interest on color images.
• Figure 4: A sample image collected from a regional Materials Recovery Facility (MRF). Materials are typically mangled, mixed, sometimes occluded and are transported through a fast conveyor system where they are sorted with automated and human-in-the-loop processes. MRF dataset collection posed challenges, with camera placement away from the conveyor belt to reduce vibrations and height limitations.
• Figure 5: Feature maps obtained from running the image sample from \ref{['fig:mrf-sample-1']} through the Mask RCNN model. The brighter regions in these feature maps indicate higher weighting in comparison to darker regions. One can interpret that these regions were used to identify plastics as a certain type and be used to distinguish from other types.
• Figure 6: Another sample image from the MRF mixed plastics conveyor belt (left). Real-world images are challenging with high variations in the number of items, their positions, size, orientation, occlusion, and impurities etc. Feature maps (right) indicate that convolutional models are adept at handling most of these variations.
• Figure 7: What do optical DNN models prioritize to determine the plastic class? (a) Object Detection using Mask RCNN (b) Model Prediction using Mask RCNN (c) Feature maps, i.e. the portions of images that contributed highly towards the classification result, indicate that shape, color and sometimes contents are used.