基于特征增强网络的交通场景图像语义分割

Semantic Segmentation of Traffic Scene Image Based on Enhancing Feature Network

针对当前交通场景图像语义分割时,因特征信息提取不足导致分割精度低的问题,提出了特征增强网络(EFN).首先将网络ResNeXt-101中卷积层5-3的输出输入到多尺度特征注意力模块中进行有针对性的特征提取;然后输出结果与卷积层2-3、卷积层3-3和卷积层4-3提取的特征图通过特征融合模块进行高效的特征融合,最后利用上采样得到图像分割结果。在数据集CamVid上的实验结果表明特征增强网络可有效提高交通场景图像语义分割的精准度。

To address the problem of low segmentation accuracy due to insufficient feature information extraction in semantic segmentation of current traffic scene images,we proposed Enhancing Feature Network(EFN).Firstly,the output of conv 5-3 layer in ResNeXt-101 is entered to the multi-scale feature attention module for targeted feature extraction,then the output is fused with the feature maps extracted from the conv 2-3、conv 3-3 and conv 4-3 layers by the feature fusion module for efficient feature fusion,and finally the image segmentation results are obtained using upsampling.Experimental results on the dataset CamVid show that EFN can effectively improves the accuracy of semantic segmentation of traffic scene image.