基于注意力机制与动态激活的空间目标检测算法

Space Target Detection Algorithm Based on Attention Mechanism and Dynamic Activation

基于图像的空间目标检测已成为保障在轨卫星运行安全的重要需求之一。已有基于深度学习的无锚框目标检测算法取得了良好进展,但是仍存在检测头结构简单、表征能力不足的问题。对此,提出了基于注意力机制与动态激活的空间目标检测算法。以无锚框目标检测算法的通用网络结构为基础,在检测头中使用基于通道与空间感知的残差注意力模块,以增强网络的特征表征能力;同时,在检测头中串联基于通道感知的动态激活模块,以提升网络在特定空间目标检测任务中的性能。在SPARK空间目标检测数据集上的实验结果表明,所提算法的AP@IoU=0.50∶0.95指标达77.1%,检测性能显著优于主流算法Faster R-CNN、YOLOv3及FCOS。此外,所提算法在训练过程中采用动态样本匹配策略,进一步提升了对小目标的检测能力。

Image-based space target detection has become one of the crucial requirements to ensure the safety of in-orbit satellites.Existing anchor-free target detection algorithms based on deep learning have achieved outstanding results.However,their detection heads have a simple structure,resulting in insufficient representation ability.To overcome this challenge,we propose a space target detection algorithm based on attention mechanism and dynamic activation.Based on the anchor-free target detection algorithm’s general network structure,the channel and spatial aware-based residual attention module is employed in the detection head to improve the network’s feature representation ability.Meanwhile,the channel aware-based dynamic activation module is connected in series with the detection head to enhance the network’s performance in a specific space target detection task.The experimental findings on the SPARK space target detection dataset demonstrate that the proposed algorithm achieves an AP@IoU=0.50:0.95 of 77.1%,and its detection performance is substantially better than the mainstream algorithms such as Faster R-CNN,YOLOv3,and FCOS.Additionally,to further enhance the detection ability for small targets,the dynamic label assignment approach is adopted in the training process.