MAFPN: a mixed local-global attention feature pyramid network for aerial object detection

Aerial object detection (AOD) aims at locating and categorizing aerial targets in remote sensing image, which has wide practical applications. Mainstream feature pyramid network (FPN) has made great progresses in multi-scale natural objects detection, which is also successfully extended into AOD. However, existing FPNs still have some shortcomings in discriminative feature extraction and fusion for aerial objects due to diverse object scales and dense small objects. To address these limitations, this letter proposes a novel mixed local-global attention FPN (MAFPN). The core components of MAFPN are cross-scale transformer (CST) module and multi-scale spatial fusion (MSSF) module, which are equipped at the feature extraction and fusion of FPN, respectively. Specifically, the CST module is devised to capture global semantic information and cross-scale similarity by exploiting self-attention in multi-scale features. The MSSF module utilizes the spatial attention maps as fusion weights, and adaptively fuses multi-scale features between adjacent two levels. Such mixed attention approach enhances the multi-scale feature learning in dense small objects. Besides, to further balance the influences of different object scales, we propose an area-guided dot-IOU (ADIOU) regression loss, which takes into account the size of objects and enables the model to prioritize the centre position of small objects. Experimental results on public datasets demonstrate the superiority of proposed method.