MF-SPNet: Dilated Convolutional Pyramid with Attention Mechanism for Semantic Segmentation of Street Scenes

In the traditional semantic segmentation of street scene, it is difficult to extract target features due to the different sizes of images, the occlusion of targets and the difficulty of recognizing small targets. Images are difficult to be semantically segmented efficiently by these models. To solve the above problems, a pyramid network using spatial attention mechanism based on atrous convolutional pyramid (AMF-SPNet) is proposed. Firstly, the feature pyramid fused with dilated convolution is used to extract the features of the target. In the original ASPP model, we change the pooling layer to atrous convolution, and change the interior of the pyramid to atrous convolution with expansion rate 1,3,6,12,18. This model can expand the receptive field of the convolution kernel without increasing the number of parameters. Secondly, the spatial attention mechanism is added to the network, so that the model can better pay attention to the street view target information. Finally, the cityspaces dataset is used for verification, and the MIoU reaches 88.3%, the MAP reaches 89.6%. Compared with other network models, the proposed model can extract the location information and spatial information of street view images more accurately, so that the accuracy of semantic segmentation is improved.