Multi-target detection of waste composition in complex environments based on an improved YOLOX-S model

Area of Science:

• Computer Vision
• Artificial Intelligence
• Environmental Science

Background:

• Accurate waste identification is vital for sustainable solid waste management.
• Challenges exist in discriminating waste categories due to incomplete features in multi-target detection.

Purpose of the Study:

• To propose an improved You Only Look Once (YOLOX-S) model for effective waste component recognition in complex scenarios.
• To enhance feature extraction and improve detection accuracy for solid waste management.

Main Methods:

• Developed an improved YOLOX-S model incorporating a convolutional block attention module, adaptive spatial feature fusion, and an efficient intersection-over-union loss function.
• Trained the model on a self-constructed dataset featuring diverse waste components and complex environmental interferences.
• Evaluated performance against original YOLO and classical models (SVM, ResNet-18, ResNet-50) on custom and public datasets.

Main Results:

• The improved YOLOX-S model achieved a mean average precision (mAP) of 85.02% on the custom dataset, a 5.32% increase over the original YOLO.
• Demonstrated reduced instances of inaccurate positioning, false detection, and missed detection.
• Outperformed classical models on a public dataset with a mAP of 94.85%.

Conclusions:

• The enhanced YOLOX-S model significantly improves waste component identification accuracy in challenging conditions.
• The model offers a robust solution for intelligent monitoring of waste, aiding in managing indiscriminate disposal and illegal dumping.
• Provides valuable decision support for emergency management in solid waste scenarios.