Data Augmentation Using Background Replacement for Automated Sorting of Littered Waste

Area of Science:

• Computer Science
• Environmental Science
• Robotics

Background:

• Automated Waste Sorting (AWS) enhances recycling efficiency.
• Littered waste in the environment poses a significant challenge.
• Current AWS systems struggle with uncollected, littered waste.

Purpose of the Study:

• Introduce BackRep, a novel method for developing waste recognizers.
• Enable identification and sorting of littered waste directly in situ.
• Improve the performance of waste recognition models for environmental applications.

Main Methods:

• Developed BackRep, a data-augmentation technique for waste recognition.
• Cropped solid waste images and superimposed them onto realistic backgrounds.
• Created a new dataset of littered waste in natural settings for training and evaluation.

Main Results:

• Waste recognizers trained on BackRep-augmented data outperformed those trained on existing datasets.
• The data-augmentation procedure demonstrated improved accuracy in identifying littered waste.
• The method shows promise for real-world deployment in diverse environments.

Conclusions:

• BackRep is a viable approach for enhancing waste recognizer development.
• The method effectively addresses the challenge of sorting littered waste.
• Supports the creation of more robust AWS systems for urban and wild environments.