The identification of waste composition based on target-detection is crucial in promoting sustainable solid waste management. However, discrimination of different solid waste categories in the presence of incomplete and insufficient feature information remains a challenge in multi-target detection. This paper proposes an improved You Only Look Once (YOLOX-S) model that enables the effective recognition of different waste components in complex environments, which enhances feature-information extraction ability regarding different dimensions by introducing a convolutional block attention module, an adaptive spatial feature fusion module, and an improved efficient intersection-over-union loss function. The improved model was trained on a self-constructed image dataset with multiple waste components and targets in various complex scenarios, including interference from similar color backgrounds, similar waste localization, and mutual waste occlusion. The experimental results showed that the improved model achieved a mean average precision (mAP) of 85.02 %, an increase of 5.32 % over the original YOLO model's mAP, and that it reduced incidents related to inaccurate positioning and false and missed detection. Moreover, the improved model outperformed classical detection models including support vector machine, RestNet-18, and RestNet-50 on a public dataset, achieving a mAP of 94.85 %. The improved model is expected to be applied to intelligent monitoring for waste components in scenarios including indiscriminate waste disposal and illegal dumping, providing decision support for emergency management.
Keywords: Deep learning; Image recognition; Multi-target detection; Waste component; YOLOX-S.
Copyright © 2024 Elsevier Ltd. All rights reserved.