一种改进变换网络的域自适应语义分割网络

语义标签的人工标注成本高,耗时长,基于域自适应的非监督语义分割是非常必要的.针对间隙大的场景或像素易限制模型训练、降低语义分割精度的问题,通过分阶段训练和可解释蒙版消除大间隙图片和像素的干扰,提出了一种改进变换网络的域自适应语义分割网络(DA-SSN).首先,针对部分源图到目标图的域间隙大、网络模型训练困难的问题,利用训练损失阈值划分大间隙的源图数据集,提出一种分阶段的变换网络训练策略,在保证小间隙源图的语义对齐基础上,提高了大间隙源图的变换质量.然后,为了进一步缩小源图中部分像素与目标图域间间隙,提出一种可解释蒙版.通过预测每个像素在源图域和目标图域之间的间隙缩小置信度,忽略对应像素的训练损失,以消除大间隙像素对其他像素语义对齐的影响,使得模型训练只关注高置信度像素的域间隙.结果表明,所提算法相比于原始的域自适应语义分割网络的分割精度更高.与其他流行算法的结果相比,所提方法获得了更高质量的语义对齐,表明了所提方法精度高的优势.

A Novel Electromagnetic Bandgap Structure for Filter

Electromagnetic bandgap （EBG） materials are periodic structures capable of prohibiting the propagation of electromagnetic waves within a certain band of frequencies. This characteristic of EBG has wide application. The structures to be studied here are mainly planar EBG materials of two dimensions, which are periodic arrays of holes etched in the ground plane of a conventional microstrip line. EBG structures are calculated with finite-difference time-domain （FDTD） method in this paper. Technique of the perfectly matched layer is used for the absorption of electromagnetic waves in FDTD. The FDTD method is programmed with the blend of C++ and Matlab languages, which makes the program both simple and fast computing. A kind of new EBG structure is brought out through a lot of experiments and analyses. A filter with wide stop-band and another filter with two stop-bands are designed.