LMF-Net: Lightweight Multidimensional Feature Fusion Network for Floating Debris Detection on Water Surfaces

Detecting floating debris on water surfaces is essential for environmental protection but remains highly challenging due to low contrast, irregular shapes, and significant noise interference in complex aquatic environments. Existing end-to-end detectors based on the DEIM framework, while efficient, struggle to preserve fine-grained spatial details and effectively suppress noise under such conditions, leading to degraded feature discriminability and frequent missed detections. To address these limitations, we propose LMF-Net, a lightweight multi-dimensional feature fusion network that enhances representation learning through three synergistic innovations. First, the Spatial-Edge Fusion module employs a dual-stage multi-path architecture to collaboratively extract local texture details and global contour features, preserving critical edge information that standard convolutions tend to smooth out. Second, the Frequency-Decomposed Hybrid Attention module decouples features into high- and low-frequency components and adaptively fuses them, effectively suppressing water ripple noise while enhancing feature discriminability. Third, the Adaptive Weighted Downsampling module replaces static pooling with content-aware feature reorganization, mitigating information loss during multi-scale pyramid construction. Experimental results on our self-built FloatDataset demonstrate that, compared to the baseline, LMF-Net reduces parameters by 1.2% and FLOPs by 0.7% while achieving significant improvements of 3.5% inAP50and 6.5% inAP50:95. On the public VisDrone2019 dataset, it outperforms state-of-the-art real-time detectors such as YOLOv8 and RT-DETR by 3.4% and 2.7% inAP50:95, respectively, demonstrating exceptional cross-domain generalization capability and superior performance.