Abstract
<title>Abstract</title> <p>With the acceleration of urbanization and industrialization, the yield of municipal solid waste has increased dramatically. Automatic waste sorting based on computer vision and deep learning has become an effective solution for environmental protection and resource recycling. However, mainstream object detection models face challenges such as insufficient multi-scale feature extraction, weak robustness to cluttered backgrounds, excessive model parameters, and high computational complexity when applied to real-world waste detection scenarios. To address these issues, this paper proposes MSF-TrashDet, a lightweight and efficient object detection network tailored for waste detection tasks. The proposed model introduces a novel Multi-Scale Feature Enhancement (MSFE) block, which integrates dynamic weight adaptation, dilated convolution, and channel-split residual connections to strengthen feature representation while maintaining computational efficiency. In addition, structured channel pruning is applied to further compress the model for practical deployment on edge devices. Extensive experiments are conducted on the public TrashNet dataset to evaluate the effectiveness of the proposed method. The experimental results demonstrate that MSF-TrashDet achieves 91.91% mAP50 with only 1.93 M parameters and 5.00 GFLOPs, which outperforms several state-of-the-art lightweight detectors including YOLOv5n, YOLOv8n, and YOLOv11n in terms of accuracy-efficiency trade-off. Ablation experiments verify the effectiveness of each key component in the proposed network. The MSF-TrashDet model provides a practical, reliable, and efficient solution for automatic waste detection and intelligent recycling systems.</p>