Research On Image Semantic Segmentation Method Based On Improved DeepLabV3+

In the task of image semantic segmentation by DeepLabV3+ algorithm,a large amount of detail information of features is lost due to the neglect of the difference of the importance degree of features in different levels of feature graphs,which leads to poor segmentation effect.Aiming at the above problems,an image semantic segmentation method based on improved DeepLabV3+ is proposed.First,Xception model is used as the backbone network,in which two low-level features were extracted simultaneously as the feature input of the decoder to add feature information.Secondly,attention mechanism is introduced,channel attention is used to weight the feature images obtained from the trunk network,and then it is fused with the feature images processed by the ASPP module to obtain rich context information and advanced features.Spatial attention was used to fuse the two low-level features and the convolutional high-level features respectively to filter a lot of background information and highlight feature points.At the same time,the depthwise separable convolution is substituted for the void convolution to reduce the parameters and improve the calculation speed.Finally,Focal Loss Loss Function is used to reduce the Loss of feature information by reducing the internal weighting.The mean Intersection over Union(mIoU),Mean Pixel Accuracy(MPA)and loss value were used as evaluation indexes.The m Io U values were 84.44% and 75.87% on PASCAL VOC 2012 and Cityscapes dataset,respectively.Compared with PSPNET,DANET and SANET algorithms,the segmentation accuracy is improved by 5.88%,3.48% and 1.24%,respectively.Experimental results show that the proposed method can effectively improve the accuracy of feature extraction and improve the final segmentation effect.There are 32 figures,11 tables and 50 references in this thesis.