Fine-tuning DETR: Toward holistic process in plastic waste sorting system

Area of Science:

• Environmental Science and Engineering
• Computer Science and Artificial Intelligence
• Materials Science

Background:

• Global plastic waste generation is a critical environmental issue, with millions of tons accumulating annually in ecosystems.
• Current waste management technologies are often outdated, posing risks to worker health and environmental safety.
• Low recycling rates exacerbate the plastic waste crisis, necessitating innovative solutions.

Purpose of the Study:

• To evaluate the efficacy of fine-tuning the DETR (Detection Transformer) model for automated plastic waste sorting.
• To analyze the performance of AI-driven plastic categorization against traditional methods.
• To assess the potential drawbacks and applicability of Transformer-based algorithms in industrial waste management.

Main Methods:

• Fine-tuning a DETR (Detection Transformer) model using an industrial plastic waste dataset.
• Comparative evaluation of the fine-tuned DETR model against other candidate algorithms.
• Assessment of key performance indicators including accuracy (mAP), processing speed (FPS), and computational cost (GFLOPs).

Main Results:

• The fine-tuned DETR model achieved superior performance with an accuracy of 25.1 mAP and a processing speed of 28 FPS.
• The model demonstrated a competitive computational cost of 86 GFLOPs.
• Fine-tuned DETR exhibited autonomous operational capabilities, reducing the need for human intervention.

Conclusions:

• Fine-tuned DETR is highly suitable for plastic waste categorization, offering significant improvements over existing methods.
• Transformer-based algorithms show substantial potential for large-scale, autonomous plastic waste sorting systems.
• Modernizing waste treatment systems with AI is crucial for environmental protection and worker safety.