空间和通道注意力多级别特征网络图像语义分割

Channel and Spatial Attention Multi-level Features Fusion for Image Semantic Segmentation

图像语义分割模型在提取特征的过程中由于不断的下采样使得空间位置丢失,分割效果较差,针对该问题,提出了融合空间和通道注意力多级别特征来构造学习网络的方法。首先针对具有高级别特征的语义信息路径引入了通道注意力模块,在预训练模型Resnet101提取的特征图上,显式地建模通道之间的相互依存关系,确定每层特征图上需要重点关注的内容,协助完成目标识别任务;其次针对具有低级别特征的空间信息路径引入空间注意力模块,在保留了丰富空间信息的特征图上提取空间注意力矩阵,并将提取的空间注意力矩阵作用于语义信息路径的相应特征图上,以确定需要重点关注的位置,协助完成目标定位任务。最后在CamVid数据集上进行实验,所提方法具有良好表现。

During the process of extracting features,the image semantic segmentation model loses the spatial position due to continuous downsampling,and the segmentation effect is poor.To solve this problem,a method of fusing spatial and channel attention multi-level features is proposed.First,the channel attention module is introduced for the semantic information path with high-level features.On the feature map extracted by the pre-training model Resnet101,the interdependence between the channels is explicitly modeled,and the focus of attention is determined on each layer of the feature map,in which the content assists in completing the task of target identification;secondly,the spatial attention module is introduced for the spatial information path with low-level features,and the spatial attention matrix is extracted from the feature map that retains the rich spatial information,and the extracted spatial attention matrix is used on the corresponding feature map of the semantic information path to determine the location that needs to be focused on,and to assist in the completion of the target positioning task.Finally,the experiments on the CamVid dataset are proved to have good performance.