SwinConvNeXt: a fused deep learning architecture for Real-time garbage image classification

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Inadequate waste management presents a global challenge, with conventional segregation methods being inefficient and costly.
• Existing deep learning models struggle with accuracy and computational efficiency in waste classification, especially for visually similar items.

Purpose of the Study:

• To develop a sophisticated and sustainable waste management system using advanced computer vision and deep learning.
• To overcome the limitations of current deep learning models in accurately classifying diverse waste types.

Main Methods:

• Proposed a novel deep learning model combining an enhanced Swin Transformer and an improved ConvNeXt with a spatial attention mechanism.
• Utilized hierarchical feature extraction and a shifting window mechanism for global feature extraction, and optimized ConvNeXt for local feature extraction.

Main Results:

• Achieved 98.97% accuracy, 98.42% precision, and 98.61% recall on the Garbage Classification dataset.
• The proposed model demonstrated superior performance over state-of-the-art deep learning models due to its lightweight design and low computational requirements.

Conclusions:

• The developed deep learning model offers an efficient and accurate solution for real-time waste classification and recycling.
• The model's effectiveness in discerning fine-grained details of waste items paves the way for improved waste management strategies.