Scale-Aware Anchor-Free Object Detection via Curriculum Learning for Remote Sensing Images

Accurate detection of multi-class instance objects in remote sensing images (RSIs) is a fundamental but challenging task the field aviation and satellite image processing, which plays crucial role wide range practical applications. Compared with natural image-based object task, RSIs-based still faces two main challenges: 1) The often present large variations size, they are densely arranged given input images; 2) Complex background distributions around tend to cause boundary blurring, making it difficult distinguish from background, resulting undesired feature learning interference. In this paper, address above challenges, we propose novel RSI anchor-free framework that consists key components: cross-channel pyramid network (CFPN) multiple foreground-attentive heads (FDHs). First, an baseline detector CFPN structure developed extract features different convolutional layers incorporates these multi-scale through parameterized processes, semantic relations across scales levels. Next, each FDH designed predict attention map enhance foreground region RSIs. Further, under scale-aware structure, design curriculum-style optimization objective dynamically reweights training instances during current epoch, enabling receive relatively easy match its ability. Experimental results on three public datasets demonstrate our method outperforms existing methods.