基于多级特征级联的遥感图像实时语义分割

Real-time semantic segmentation of high-resolution remote sensing image based on multi-level feature cascade

针对遥感图像语义分割中存在的分割耗时长、分割小目标不准确的问题,提出基于多级特征级联的高分辨率遥感图像快速语义分割模型(multi-level feature cascade network, MFCNet)。该模型主要由特征编码、特征融合以及目标细化3部分组成。特征编码对输入的不同分辨率图像用不同量级主干网络进行特征提取,由于低分辨率图像分辨率较低,使用重量级的主干网络在增加较少参数的情况下可以获取丰富的语义信息,而中、高分辨图像分辨率较大,使用轻量级主干网络既减少参数量又可获取全局信息。中等和低分辨率的编码使用权重和计算共享的方式,进一步减少模型参数,降低计算复杂性。特征融合对来自不同分支的特征进行融合,以获取不同尺度的信息。目标细化采用残差校正对融合后的特征和编码部分的特征进行融合校正,以恢复图像的空间细节信息,使分割更加准确。该模型可以端到端的方式有效地工作,试验验证所提模型在遥感图像语义分割中的有效性,在模型复杂性和精度上取得较好的平衡。

Aiming at the problems of long segmentation time and inaccurate segmentation of small targets in remote sensing image semantic segmentation, a fast semantic segmentation model of high-resolution remote sensing image based on multi-level feature cascade network(MFCNet) was proposed. The model was mainly composed of feature encoding, feature fusion and target refinement. Feature encoding extracted the input images feature of different resolutions and used different backbone networks. Due to the lower resolution of low-resolution images, heavy-weight backbone networks were used to obtain rich semantic information with fewer parameters. For medium and high-resolution images, lightweight backbone network was used to reduce the amount of parameters and obtain global information. While medium and low-resolution encoding used the way of weights and calculation sharing to further reduce model parameters and computational complexity. The feature fusion section fused features from different branches to obtain information at different scales. The target refinement used residual to correction the fused features and the features of the coded part to restore the spatial detail information of the image, making the segmentation more accurate. And the entire model worked efficiently in an end-to-end manner. The experimental verified the validity of the model in semantic segmentation of remote sensing images, and achieved a good balance between model complexity and accuracy.