| Semantic segmentation technology is a challenging task in computer vision,and it plays a crucial role in image understanding.Its objective is to accurately segment objects and assign semantic labels to the segmented images,which provides valuable information to users in an efficient manner.This article explores the potential of deep learning models by starting with the most popular semantic segmentation network for street view images.By incorporating attention mechanism modules,replacing backbone networks,modifying loss structures,and implementing other techniques,the image segmentation algorithms have been enhanced,paving the way for further improvements in the area of street view image semantic segmentation.The main tasks are as follows:(1)U-Net semantic segmentation with ECA attention mechanism.This chapter proposes a new U-Net network structure based on Transformer.During the fusion process,a super strong channel attention module is introduced to enhance the discriminative ability of feature representation for semantic segmentation of street view images.The hole convolution is introduced into the overall network structure to expand the Receptive field range of the street scene image without changing the size of the convolution core,so as to maximize the network performance.(2)New Deeplab v3 semantic segmentation based on auxiliary loss.This chapter ensures that the model can significantly reduce computational complexity when facing the same output by integrating a pyramid pooling head structure with deep separable convolutions.By setting weights in stages and setting auxiliary heads as auxiliary losses,guidance can be provided for shallow network weights.The improved model sets the sampling mechanism of difficult samples(OHEM),introduces void convolution and multi task loss function,and improves the segmentation effect of the model from many aspects.(3)Deeplab v3+ semantic segmentation with the Focal Loss function.This chapter introduces the Focal Loss function to reduce the training weight of a large number of simple street view samples,effectively solving the problem of severe imbalance in the proportion of positive and negative samples.The use of the SE attention mechanism module enhances the model’s attention to a certain type of feature channel,suppresses its attention to other types of feature channels,and improves the model’s segmentation performance. |
