Remote Sensing Object Detection based on Attention and Feature Fusion

In addressing the challenges of significant scale variation and high resemblance between the foreground and background, we present a one-stage remote sensing object detection strategy in this paper. Specifically, we have developed a multi-scale information mining module, integrated into the backbone network, to enhance feature representation capabilities and address the issue of vast scale variations in objects. Subsequently, we utilize a deep and shallow feature fusion module to harmonize shallow and deep features. This module not only effectively detects multi-scale objects but also improves the precision detection of smaller objects. To alleviate the problem of foreground and background similarities, we have incorporated a dual path attention mechanism into our feature pyramid networks. This adaptation enables the network to focus more intensively on object information. Comprehensive experiments on two distinct remote sensing object detection datasets, namely, DIOR and NWPU VHR-10, validate the efficacy of our proposed method. Furthermore, our approach demonstrates its superiority over current state-of-the-art methodologies, improving the baseline method by 3.3% and 15% respectively.