Research On Image Semantic Segmentation Based On Improved DeepLabV3+ Network Structure

Image semantic segmentation has become one of the very hot research topics in the field of artificial intelligence,computer vision and deep learning.This detailed pixel-level understanding of image semantic segmentation is crucial for many AI-based systems to enable a comprehensive understanding of scenarios including medical image segmentation,robotics,self-driving cars,etc.Therefore,the study of image semantic segmentation is very valuable.In this paper,the PASCAL VOC2012 dataset is used as the research object,and the DeepLabV3+ network framework is used to achieve semantic segmentation of images.With the objective of improving the image segmentation accuracy,the DeepLabV3+ network is improved based on the attention mechanism,cross-layer feature fusion and cosine domain loss function.The main work accomplished in this paper is as follows:(1)To obtain higher segmentation accuracy,two structures,Xception＿65 and Xception＿71,are used to complete the semantic segmentation training of DeepLabV3+ network respectively,and the segmentation accuracy is evaluated according to the evaluation index MIoU of image semantic segmentation.The results show that Xception＿71 as the backbone network of DeepLabV3+ has better semantic segmentation on edges and details.(2)In order to obtain the full range of attentional feature information,a DeepLabV3+image semantic segmentation algorithm based on the attention mechanism is proposed.Firstly,the Convolutional Block Attention Module(CBAM)is introduced into the DeepLabV3+network,and the CBAM module is optimized so that the original feature maps and CAM results are summed and input to the SAM.Secondly,to address the problem of pixel discontinuity caused by direct upsampling of the decoding area by a factor of 4,the shallow features of different layers are fused to the deep features across layers and then fine-tuned for upsampling to achieve optimal segmentation.Tests show that the proposed DeepLabV3+ algorithm based on the attention mechanism achieves higher segmentation accuracy.(3)To ensure intra-class densities and inter-class dispersions,an Attention-DeepLabV3+image semantic segmentation method based on the cosine-domain loss function is proposed.The method uses a cosine-domain loss function instead of the Softmax loss function,incorporates an angle margin in the feature space and a penalty mechanism on the cosine value.The model is tested to verify the soundness of the model,and the results conclude that better segmentation results with more detail and accuracy can be obtained by the cosine domain loss function.