Source-linked AI summary
Waste detection in Pomerania: non-profit project for detecting waste in environment
Sylwia Majchrowska, Agnieszka Mikołajczyk, Maria Ferlin, Zuzanna Klawikowska, Marta A. Plantykow, Arkadiusz Kwasigroch, Karol Majek
TL;DR
Waste pollution and difficult, inconsistent segregation create a need for more capable automated litter processing. The paper develops an open-source framework using diverse waste datasets, seven categories, and separate detection and classification networks. It achieves 66.4% AP@0.5 for EfficientDet-D2 detection and up to 75% accuracy for EfficientNet-B2 classification, while small-object detection and imbalanced data remain limitations.
Problem
Waste management is hindered by complex segregation rules, manual plastic separation, limited diverse datasets, and few attempts to classify litter into recognized recyclable classes.
Method
The paper builds an open-source two-stage framework that detects litter with one neural network and classifies it into seven categories with another, using mixed public datasets and pseudo-labeling.
Results
66.4% AP@0.5 was achieved for EfficientDet-D2 detection, and up to 75% accuracy was achieved for EfficientNet-B2 litter classification.
Takeaways & Limitations
The framework supports universal litter detection and classification across waste observed in natural, urban, indoor, and underwater environments.
Takeaways & Limitations
Small-litter recognition remains challenging, while classification performance is constrained by dataset imbalance and limited labeled examples for some categories.
Abstract
from arXiv · showhide
Waste pollution is one of the most significant environmental issues in the modern world. The importance of recycling is well known, either for economic or ecological reasons, and the industry demands high efficiency. Our team conducted comprehensive research on Artificial Intelligence usage in waste detection and classification to fight the world's waste pollution problem. As a result an open-source framework that enables the detection and classification of litter was developed. The final pipeline consists of two neural networks: one that detects litter and a second responsible for litter classification. Waste is classified into seven categories: bio, glass, metal and plastic, non-recyclable, other, paper and unknown. Our approach achieves up to 70% of average precision in waste detection and around 75% of classification accuracy on the test dataset. The code used in the studies is publicly available online.
1 Introduction
Waste pollution and inconsistent segregation practices motivate automated litter sorting. The paper responds with an open-source framework combining diverse datasets, seven waste categories, and separate detection and classification networks.
- Waste production has reached about 2 billion tons annually, while plastic production exceeds 300 million tons and ocean garbage is projected to surpass sea creatures within 30 years.
- Complex, non-unified segregation guidelines and the energy costs of manual plastic separation make waste management difficult.
- Existing smart-bin systems commonly classify one object at a time against a clear background, limiting their direct applicability to varied environments.
- The proposed framework mixes publicly available datasets, defines seven sorting categories, and uses separate detector and classifier neural networks.
- The paper reviews existing waste datasets and presents training, evaluation, and conclusions for its detection-and-classification framework.
2 Related works
Related work covers deep learning architectures, object-detection paradigms, and waste datasets. It highlights the limited prior use of diverse data for detection and classification into well-recognized recyclable classes.
- Deep learning architectures influence image-recognition accuracy, latency, and computational requirements, shifting emphasis toward neural-network design.
- CNN research progressed from AlexNet and VGG to residual, multi-branch, densely connected, cardinality-based, and EfficientNet architectures.
- EfficientNet scales network width, depth, and resolution, while EfficientNetv2 targets faster training and improved parameter efficiency.
- Object detection locates axis-aligned bounding boxes and assigns classes through either one-stage prediction or two-stage proposal-and-classification pipelines.
- Earlier two-stage methods generated multi-scale proposals with sliding windows before classifying them, while region-based CNN methods extracted richer proposal features.
- Single-stage SSD and YOLO detectors combine localization and classification in one step to reduce detection time.
- Few studies detect and classify litter into recognized recyclable classes, motivating a review of over ten datasets and their key statistics.
3 Two-stage framework to detect and sort litter
The framework addresses diverse waste environments and inconsistent sorting categories by separating litter localization from category classification. It combines public datasets, neural-network experiments, and pseudo-labeling to evaluate detection and seven-class waste classification.
- Framework: The pipeline first localizes litter regions, then extracts each region for a separate classifier to assign its waste category.This two-stage design separates detection from classification.
- Data: Twelve public datasets plus Google Images data were combined into detection and classification resources, including detect-waste+ with cigarette-butt images for small-object detection.The detection dataset used one Litter class, while additional category images supported classification.
- Data limitations: The classification data were highly imbalanced, dominated by metals and plastic and followed by unknown litter, with annotation errors also reported.The authors identify inconsistent annotation rules as an additional dataset problem.
- Detection results: 65.5% average precision was achieved by EfficientDet-D2 on the one-class detect-waste dataset, outperforming the tested Mask R-CNN with ResNet-50 baseline at 28.0%.EfficientDet-D2 was selected because it achieved the best evaluation results with fewer parameters and lower computing requirements than the alternatives considered.
- Classification results: Seven-class detection reduced mAP to 16.2% at IoU 0.50 and 13.0% at IoU 0.75, whereas classification accuracy reached 74.6% with a random sampler and 73% with a weighted sampler.EfficientNet-B2 exceeded ResNet-50 by over 10 percentage points in classifier comparisons, while weighted sampling produced more balanced class results.
4 Conclusions and Future Work
The paper presents an open, two-network framework for detecting and classifying litter across diverse environments, using seven sorting categories. Detection reached 66.4% AP@0.5 with EfficientDet-D2, while classification achieved up to 75% accuracy with EfficientNet-B2, but small-litter recognition remains challenging.
- Framework: The framework localizes trash and identifies its class using two separate neural networks trained on waste data from indoor, natural, urban, and underwater environments.The datasets combined publicly available waste images collected across these settings.
- Detection: 66.4% AP@0.5 was achieved for waste detection with EfficientDet-D2 after adapting three neural-network architectures and naturally augmenting the data.Qualitatively good results were also observed for the semi-labeled OpenLitterMap dataset.
- Classification: Up to 75% accuracy was achieved for litter classification with EfficientNet-B2 across seven sorting categories plus a background class.Pseudo-labeling used unlabeled data but provided only a slight classification-performance boost, associated with imbalance and limited labels for some categories.
- Future Work: Small-litter recognition remains challenging, motivating future work with alternative detectors, more balanced data, and newer EfficientNetv2 models.The authors identify improved performance as the main future direction because AI is expected to be more accurate than humans.
- Implications: The framework showed potential for waste management applications, including future robotic sorting without human intervention.The authors also connect high localization precision across varied environments with possible neural-network use for waste monitoring.