A deep convolutional neural network to simultaneously localize and recognize waste types in images.

Area of Science:

• Computer Science
• Environmental Science
• Artificial Intelligence

Background:

• Accurate waste classification is crucial for effective waste management.
• Existing methods often focus on single-label classification, which is insufficient for real-world scenarios.

Purpose of the Study:

• To develop and evaluate a benchmark for multi-label waste classification and localization using deep learning.
• To propose a novel multi-task learning architecture (MTLA) for simultaneous waste identification and localization.

Main Methods:

• A multi-task learning architecture (MTLA) based on a convolutional neural network was proposed.
• The MTLA incorporates attention modules, a multi-level feature pyramid network, and joint learning subnets.
• Novel loss functions focusing on joint optimization of classification and localization were designed.

Main Results:

• The MTLA achieved an F1 score exceeding 95.50% for multi-label waste classification.
• The model obtained an average precision score over 81.50% for waste localization.
• Performance was comparable to that of human experts, demonstrating high accuracy in waste management tasks.

Conclusions:

• The proposed MTLA effectively addresses multi-label waste classification and localization challenges.
• The MTLA shows promise as an auxiliary tool to enhance automation in waste management systems.
• Visualizations using heatmaps aid in interpreting the model's feature extraction process.