Focus-RCNet: a lightweight recyclable waste classification algorithm based on focus and knowledge distillation.

Area of Science:

• Environmental Science
• Computer Science
• Artificial Intelligence

Background:

• Global waste pollution is escalating due to improved living standards and consumption patterns.
• Automated waste classification using artificial intelligence offers a promising solution.
• Existing convolutional neural networks are often too computationally intensive for real-time embedded systems.

Purpose of the Study:

• To develop a lightweight network architecture for efficient and accurate recyclable waste image classification.
• To overcome the computational limitations of traditional deep learning models in embedded applications.

Main Methods:

• Proposed Focus-RCNet, a lightweight network inspired by MobileNetV2's structure, utilizing deep separable convolutions.
• Integrated a Focus module for feature dimensionality reduction while preserving key information.
• Employed the SimAM attention mechanism to enhance feature focus with minimal parameters.
• Applied knowledge distillation to further compress model parameters.

Main Results:

• The Focus-RCNet model achieved an accuracy of 92% on the TrashNet dataset.
• Demonstrated high deployment mobility, making it suitable for real-time applications.
• Successfully reduced model complexity and parameter count compared to traditional networks.

Conclusions:

• Focus-RCNet offers an effective and efficient solution for automated recyclable waste classification.
• The model's lightweight design and high accuracy make it ideal for real-time embedded systems.
• This approach contributes to addressing the growing challenge of waste management through AI innovation.