融合深层差异特征的RGB-T巢式语义分割网络

Nested semantic segmentation network fusing deep difference features

针对现存可见光—红外(RGB-T)图像语义分割模型分割性能不高的问题,提出一种基于深层差异特征互补融合的巢式分割网络。具体来说,网络的编码和解码部分通过多级稠密中间路径相连形成一个嵌套形式的结构,编码器的深浅特征通过多级路径供解码器实现密集的多尺度特征复用,另一方面多模态深层特征通过特征差异性融合策略增强其语义表达能力。实验结果表明,所提网络在MFNet数据集上实现了65.8%的平均准确率和54.7%的平均交并比,与其他先进RGB-T分割模型相比,具有更优越的分割能力。

Considering the existing visible-infrared image(RGB-T)semantic segmentation models have limitations in segmentation performance,this paper proposed a nested semantic segmentation network fusing deep difference features.Specifically,it connected the encoding part and the decoding part of the network by a multi-level dense intermediate path to form a nested structure,and encoder features at various levels achieved densely repeated utilization via multi-stage path while the multi-modal deep feature enhanced its semantic expressiveness by the feature differential fusion strategy.The comparison experiments show that the proposed network achieves an average accuracy of 65.8%and an average intersection over union of 54.7%on the MFNet dataset.Compared with other state-of-the-art RGB-T segmentation models,it has better segmentation ability.