Applying a deep residual network coupling with transfer learning for recyclable waste sorting.

Area of Science:

• Artificial Intelligence
• Environmental Science
• Computer Vision

Background:

• Effective recyclable waste sorting is crucial for circular economy development and achieving carbon neutrality.
• Current methods require intelligent and efficient solutions for improved waste management.

Purpose of the Study:

• To develop an intelligent and efficient deep learning-based method for recyclable waste classification.
• To propose and evaluate various ResNet architectures (RWNet models) for this task.

Main Methods:

• Utilized transfer learning with ResNet structures (ResNet-18 to ResNet-152) to create RWNet models.
• Implemented cyclical learning rate and data augmentation to enhance model performance.
• Evaluated models using accuracy, precision, recall, F1 score, and ROC analysis.

Main Results:

• RWNet models achieved high performance, with accuracy nearing 88.8% (RWNet-152).
• RWNet-101 and RWNet-152 demonstrated the highest weighted average precision (89.9%), recall (88.8%), and F1 score (88.9%).
• Area under the ROC curve (AUC) exceeded 0.9 for most waste types, indicating robust classification capabilities.

Conclusions:

• RWNet models show strong performance in automatically sorting most recyclable waste types.
• The developed method offers a promising solution for automated waste management and contributes to circular economy initiatives.