LPW-Net: A Lightweight Approach for Detecting Floating Debris on River and Lake Surfaces

To address the challenges of small target sizes, diverse shapes, and complex environments, which lead to low detection accuracy and large model sizes, we propose LPW-Net, a lightweight method for detecting floating objects on river and lake surfaces. We introduce a specially designed small object detection head to enhance accuracy for small targets. To improve the backbone network’s feature extraction, we design an efficient module, C2f_ODSLK, based on depthwise separable convolutions. Additionally, we incorporate GSConv convolution and a new lightweight feature fusion module, C2f_PGA, in the Neck section, ensuring high detection accuracy while reducing model size. Experimental results show that LPW-Net reduces parameter count by 16.4% and model size by 12.7% compared to YOLOv8n, achieving mAP50 and mAP50-95 improvements of 3.2% and 2.6% on our dataset, outperforming other algorithms. The model also demonstrates significant competitiveness on the VisDrone2019 dataset.