Related Experiment Videos
Adaptive multi-modal dynamic graph neural networks: enhancing structure modeling and discriminability for waste
Yuhang Yang1, Yuanqing Luo2, Yingyu Yang1
1School of Environmental and Chemical Engineering, Shenyang University of Technology, Shenyang 110870, China.
Waste Management (New York, N.Y.)
|August 11, 2026
Summary
This study introduces an advanced waste classification system using a novel dynamic graph neural network. The method achieves over 99% accuracy, offering a robust solution for automated waste sorting.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Existing waste classification methods struggle with integrating visual and semantic information.
- Current approaches often lack robustness in structural modeling for waste items.
Purpose of the Study:
- To develop an adaptive multi-modal dynamic graph neural network framework for enhanced waste classification.
- To improve the integration of visual and semantic cues and structural modeling in waste sorting.
Main Methods:
- Fusing features from ResNet50 and CLIP ViT-B/32.
- Employing an adaptive dynamic graph construction mechanism (DKNN).
- Utilizing a multi-head dynamic attention module and supervised contrastive loss.
Main Results:
- Achieved over 99% accuracy on TrashNet and a proprietary dataset.
- Significantly outperformed state-of-the-art methods in accuracy, convergence speed, and stability.
- Demonstrated practical applicability on an industrial-scale experimental platform.
Conclusions:
- The proposed framework offers an efficient and reliable solution for industrial waste sorting automation.
- The method effectively addresses limitations in current waste classification techniques.
- Validated practical applicability for real-world waste management scenarios.