边界辅助监督的轻量化卫星部件语义分割网络

Edge-Auxiliary Supervised Satellite Components Segmentation Network

空间卫星部件的准确分割对于在轨对接及维修等具有重要意义,然而太空场景的特殊性导致卫星及其部件和太空背景的像素数量差异悬殊,同时不同部件在相邻区域难以区分.本文提出边界辅助监督的轻量化卫星部件语义分割网络(edge-auxiliary supervised components segmentation network,EASCSN),经过双分支结构提取多尺度特征和全局语义信息,并以级联形式将语义信息注入到空间特征中,采用多尺度特征聚合解码器处理编码器特征并完成部件分割.此外,为了提高网络边界感知能力,增加边界监督策略辅助训练.在公开数据集的定性与定量实验证明,本方法能在实时且轻量的前提下准确完成卫星部件语义分割,其中平均交并比达到74.74%,平均准确率达到80.99%,计算量和参数量仅为6.08GFLOPS(giga floating-point operations per second)和5.45M,在NVIDIA Tesla T4 GPU上分割速度能够达到43.29帧每秒.高效的卫星部件识别具有辅助感知目标卫星结构和完成空间任务的应用价值,并具备推广星上计算机的潜力.

Precise segmentation of satellite components is key to RPO(rendezvous and proximity operations)and OOS(on orbit servicing),while harsh space environment and compact layout of components hinder fine-grained pixel-wise recognition.An edge-auxiliary supervised components segmentation network(EASCSN)is proposed to tackle these problems.First,a two-branch encoder is designed where pyramidal spatial features and global semantic features are fused with gated semantic injection module in a cascade manner.Second,a slim but strong decoder with convolution-free feature aggregation module is elaborately designed so that high-quality parts of multi-scale features are distilled and aggregated.Meanwhile,an auxiliary edge-supervised strategy is adopted during training for sharper prediction in the edge regions.Massive experiments demonstrate the superiority of the proposed EASCSN.With compact model size and low computation cost,EASCSN can achieve a new state-of-the-art speed-accuracy trade-off.Specifically,on a single Tesla T4 GPU,EASCSN yields 74.74%mIoU and 80.99%mAcc at 43.29 FPS on UESD test set.Efficient satellite components recognization helps perceive structure of target satellites and achieve space intelligent control.There is potential value of being further deployed to spaceborne platforms.