MSWNet: A visual deep machine learning method adopting transfer learning based upon ResNet 50 for municipal solid...

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Global municipal solid waste (MSW) is increasing, necessitating intelligent and efficient sorting methods.
• Traditional manual and semi-mechanical sorting methods are inefficient, resource-intensive, and can accelerate virus transmission.
• The diverse composition of MSW complicates traditional sorting processes, leading to low efficiencies.

Purpose of the Study:

• To introduce and evaluate MSWNet, a deep learning model, for intelligent MSW sorting.
• To enhance MSW classification accuracy and efficiency using transfer learning and cyclical learning rates.
• To improve the transparency and accountability of deep learning models in MSW sorting through visualization.

Main Methods:

• Application of MSWNet, a ResNet-50 model with transfer learning, for MSW classification.
• Utilizing a cyclical learning rate strategy to optimize model training.
• Implementing visualization techniques to enhance model interpretability.

Main Results:

• Transfer learning reduced training time (741s to 598.5s) and improved accuracy (88.50% to 93.50%).
• MSWNet achieved high performance in MSW classification: 93.50% sensitivity, 93.40% precision, 93.40% F1-score, 93.50% accuracy, and 92.00% AUC.
• The study demonstrated the effectiveness of deep learning for MSW classification and data dimensionality reduction.

Conclusions:

• MSWNet offers a significant advancement in intelligent MSW sorting.
• Deep learning models, with appropriate training strategies like cyclical learning rates, can achieve high classification accuracy.
• The findings provide a reference for developing deep learning models for MSW classification and data processing.