Abstract
Remote sensing images object detection as a research hots-pot in recent years, its detection effect and inference speed are attracting much attention. Small receptive field often lead to object classification errors because of the similarity features between different categories. In addition, the large size of remote sensing images leads to slow inference speed. To address above problems, this paper proposes a single-stage rotated object detector RS-LMF2. Firstly, ResNet-Dil module is used to increase the receiver field of the model, and then the Augment-FPN module is used to merge the feature information between the bottom layer and the top layer to obtain prior knowledge, so that the model can capture enough background information in the remote sensing objects to increase the detection effect of the model. In order to improve inference speed, this paper designs the refined sparse module, which not only reduces the number of initial settings of anchor, but also uses multiple convolutions to obtain the angle information of the objects, so that the horizontal box is gradually regressed into a rotated box to improve the inference speed. RS-LMF2 achieves excellent results in two datasets, i.e., DOTA (79.0% mAP, 22.3 FPS), and UCAS-AOD (90.8% mAP, 39.2 FPS) on an NVIDIA 3090 GPU.