Zero-shot and few-shot multimodal plastic waste classification with vision-language models

Area of Science:

• Materials Science
• Computer Science
• Environmental Science

Background:

• The construction sector generates significant plastic waste, necessitating efficient recycling processes.
• Accurate classification of plastic waste is crucial for value retention during recycling.
• Current supervised deep learning models require extensive labeled data, limiting scalability and generalizability.

Purpose of the Study:

• To explore the application of Vision-Language Models (VLMs) for classifying construction and demolition plastic waste by resin type.
• To evaluate the efficacy of VLMs in zero-shot and few-shot learning scenarios for plastic waste classification.
• To compare VLM performance against fully supervised methods regarding accuracy, scalability, and data efficiency.

Main Methods:

• Utilized advanced VLMs for zero-shot classification of plastic waste using natural language descriptions.
• Integrated image and textual modalities within multimodal few-shot learning frameworks.
• Conducted comprehensive experiments comparing VLM performance with data-intensive, fully supervised baselines.

Main Results:

• VLMs demonstrated effective classification of end-of-life plastics with minimal to no training data.
• Achieved 70.15% accuracy in zero-shot classification using VLMs.
• Improved accuracy to 85.07% with multimodal few-shot learning, highlighting enhanced data efficiency and scalability.

Conclusions:

• VLMs offer a promising, data-efficient approach for classifying construction and demolition plastic waste.
• The zero-shot and few-shot learning capabilities of VLMs significantly enhance scalability for plastic waste management.
• VLMs present a viable alternative to traditional supervised methods, addressing challenges in data acquisition and model generalizability.