MARFPNet: Multiattention and Adaptive Reparameterized Feature Pyramid Network for Small Target Detection on Water Surfaces

Abstract

The images captured by unmanned aerial vehicles (UAVs) are often limited in scale and feature information, making it challenging for current detection algorithms to learn the features of objects effectively. This limitation hampers accurate identification of small objects on water surfaces. We introduce a multiattention and adaptive reparameterized feature pyramid network for small target detection on water surfaces (MARFPNet) to tackle this issue. First, to address the loss of small object features during extraction, we improved the attention mechanism based on the characteristics of small objects and proposed a multiattention module, integrating it into the feature extraction process. Second, to address the semantic information of small objects being retained mostly in shallow feature maps and not fully utilized, we introduced an adaptive reparameterized generalized feature pyramid network (Adaptive_RepGFPN). This module reorganizes features, expands the fusion scale, and incorporates adaptive weighting in the concat operation. Third, to overcome the challenge of ineffective restoration of feature map information by upsampling, we introduce the Dysample. Finally, to address the problem of the loss function being sensitive to scale changes, we propose the normalized Wasserstein distance (NWD) loss function to reduce the sudden drop in loss due to scale changes. We conducted experiments on VisDrone, SeaDronsSee, and the self-build dataset. MARFPNet showed higher accuracy compared to other detection algorithms. Notably, on the self-build dataset, mAP50 and mAP50:95 improved by 9.1% and 3.5% over the baseline network. This demonstrates MARFPNet’s effectiveness and suitability for detecting small targets in drone aerial photography on water surfaces.