Waste sorting models routinely exceed 95% accuracy on benchmark datasets, yet they ignore a critical real-world problem: contamination in recyclable waste. A contaminated item—such as a pizza box with grease—should be sent to garbage, not recycling, but conventional classifiers route it incorrectly. A new two-stage deep convolutional neural network called EcoBin, developed by Kumar and Raghav Senthil (published on arXiv, June 2026), targets this blind spot by adding a dedicated contamination-checking stage.
How EcoBin Works
EcoBin consists of two sequential neural networks. The first stage is a base waste classifier built on an EfficientNetV2-S backbone. It assigns each of thirty waste categories to one of four disposal pathways (e.g., recycling, compost, landfill). The second stage is a contamination classifier that inspects any item routed toward recycling and overrides the decision to garbage when contamination is detected.
Because no public dataset of contaminated recyclables exists, the researchers synthesized one. They used a U2-Net model to segment images of clean recyclable objects and composited realistic contamination textures onto their surfaces. This synthetic dataset trains the contamination classifier, which achieves a 0.99 ROC-AUC in distinguishing clean from contaminated items.
Performance Metrics
The base classifier achieves 87.42% test accuracy and a 96.13% pathway-adjusted accuracy (meaning it assigns most items to the correct disposal pathway even if the exact category is wrong). However, on a test set of contaminated recyclables, the base classifier alone routes only 1 of 25 items correctly—it does not detect contamination. With the full EcoBin pipeline, that number jumps to 24 of 25 items correctly routed. A McNemar's test confirms the contamination stage contributes a statistically significant improvement (p < 0.001).
| Metric | Base Classifier Alone | Full EcoBin Pipeline |
|---|---|---|
| Test accuracy (30 categories) | 87.42% | — |
| Pathway-adjusted accuracy | 96.13% | — |
| Contaminated recyclables correct routing | 1/25 (4%) | 24/25 (96%) |
| Contamination detection (ROC-AUC) | — | 0.99 |
| McNemar's test p-value | — | p < 0.001 |
Implications for Supply Chain and Logistics
While the source paper does not specify commercial deployment, EcoBin's architecture has clear relevance for waste management and recycling supply chains. Contamination is a persistent problem: a single dirty item can spoil an entire batch of recyclables, increasing processing costs and reducing the quality of recycled material. An AI system that reliably flags contamination at the bin or collection point could reduce sorting errors, improve material purity, and lower operational costs for recycling facilities. The two-stage design—classify first, then check for contamination only on recyclable-bound items—is computationally efficient, as the contamination stage operates only on a subset of items.
The use of synthetic data generation (U2-Net segmentation + texture compositing) is another key enabler, because real contaminated waste images are rare and difficult to collect. This approach could be adapted by logistics technology vendors building AI for waste audit systems, reverse logistics, or circular economy platforms.
Technical Stack
- Base classifier backbone: EfficientNetV2-S
- Contamination classifier: trained on synthetic data generated via U2-Net segmentation
- Training data: custom dataset of 30 waste categories with synthetic contamination
- Evaluation metrics: test accuracy, pathway-adjusted accuracy, ROC-AUC, McNemar's test
Competitive Context
Most existing waste classification models, such as those used in smart bins or recycling robots, do not account for contamination. EcoBin explicitly addresses this gap. The paper compares its pipeline to the base classifier alone, showing a dramatic improvement on contaminated items. No commercial product is named, but the approach is directly applicable to waste sorting systems sold by companies like AMP Robotics, Bulk Handling Systems, or Tomra.
The paper is available on arXiv under a CC BY-SA 4.0 license, allowing reuse and adaptation. Further research could validate the model on real-world contaminated waste streams and integrate it into IoT-enabled bins or conveyor-belt sorting systems.