Dynamic Pyramid Attention Networks for multi-orientation object detection

Dynamic Pyramid Attention Networks for multi-orientation object detection

The objects in remote sensing images often appeared in any direction, and thus multi-orientation object detection has received considerable attention. However, most existing oriented object detection methods rely on increasing the network layers, which wastes many computing resources while only bringing a slight improvement. We find that a few pixels around the convolution kernel participate in the calculation when extracting image features in the convolution network. If we can incorporate the global information into the feature map, the model’s performance will be significantly improved. In this paper, we proposed the dynamic pyramid attention network (DPANet) for remote sensing images, which consists of the self-attention feature pyramid network (SAFPN) and the dynamic feature map selection module (DFMS). The SAFPN employs the self-attention mechanism to learn the correlation between each pixel value and the global pixel in different feature layers by downsampling the upper feature layers to the lower one. Furthermore, the DFMS module dynamically selects feature maps to further expand the receptive field by weighing the effectiveness of different feature layers and reducing the interference of unnecessary feature maps. The remote sensing datasets HRSC2016, UCAS_AOD and NWPU_VHR are used to evaluate the performance of DPANet and the experiment results show that the proposed network outperforms the benchmark models significantly.