基于改进U-Net的街景图像语义分割方法

为提升多尺度目标的分割效果，增强特征提取能力，提出了一种基于双重注意力机制的改进U-Net街景图像语义分割方法。在U-Net编码阶段的第5个卷积块之后，添加特征金字塔注意力模块，提取多尺度特征，融合上下文信息，增强目标语义特征。在解码阶段不再采用U-Net的特征拼接方法，而是设计了一个空间域-通道域联合注意力模块，接收来自跳跃连接的低层特征图和来自前一个注意力模块的高层特征图。在Cityscapes数据集上的实验结果表明，引入的注意力模块可有效提升街景图像分割精度，与PSPNet、FCN等方法相比，分割性能指标mIoU提升了2.0%～9.6%。

Semantic segmentation method of street view image based on improved U-Net

An improved U-Net street image semantic segmentation method based on a dual attention mechanism is proposed to improve the segmentation effect of multi-scale targets and enhance the feature extraction ability.After the fifth convolutional block in the U-Net encoding stage, the feature pyramid attention module is added to extract multi-scale features, fuse contextual information, and enhance the target semantic features.Instead of using the feature stitching method of U-Net in the decoding stage, a joint spatial domain-channel domain attention module is designed to receive the low-level feature maps from the jump connection and the high-level feature maps from the previous attention module.Experimental results on the Cityscapes dataset show that the introduced attention module can effectively improve the street view image segmentation accuracy, and the segmentation performance metric mIoU improves by 2.0~9.6 percentage points compared with methods such as PSPNet and FCN.