Research On Image Semantic Segmentation Model Based On Improved DeepLabv3+

Image semantic segmentation is an important area of research directions in the field of computer vision,which requires accurate classification of image pixels,precise positioning of image content and allocation labels to different image areas.Technology of deep learning-based image semantic segmentation has achieved excellent results in autonomous driving,medical imaging and other fields.However,existing methods still face challenges due to the limitations of deep learning,as well as the complexity and diversity of practical scenes,such as complex and diverse targets,mutual occlusion,difficult to balance reasoning speed and segmentation accuracy.Therefore,this paper proposes the following two image semantic segmentation models:(1)In order to solve the problem of unsatisfactory segmentation of some small target objects in DeepLabv3+ model and the loss of some details caused by only one scale coding features in the decoding stage,a semantic segmentation model based on attention mechanism and strip pooling is proposed.Firstly,the strip attention refinement module is introduced after the backbone network to strengthen the learning of important features and improve the perception ability of small target information;Secondly,the improved mixed strip pooling module is used to reconstruct the ASPP module to enhance contextual information relevance and generate richer high-level semantic features.Finally,the spatial attention feature fusion module is designed in the decoding stage to integrate the multi-scale features extracted from the backbone network to enrich the low-level semantic features.Experimental results on Cityscapes and PASCAL VOC 2012 datasets show that the model can improve the ability of encoders to extract features,improve the perception of small targets and alleviate the problem of loss of detail information.and MIo U improved by 2.31% and 3.15% compared with the original model.(2)In the field of semantic segmentation,DeepLabv3+ model has met the requirements of high precision segmentation results.However,the overall structure parameters of this model are large,which cannot well balance the semantic segmentation rate and segmentation accuracy.Therefore,the DeepLabv3+ image semantic segmentation model based on Mobile Netv2 and attention mechanism is proposed.First,the Mobile Netv2 embedded in the improved SE attention module is used as the backbone network of the model to reduce the number of model parameters,improve the attention of important targets,and capture richer features;Secondly,the improved depth separable convolution is introduced in the ASPP module,and the void rate of void convolution is recombined,so as to reduce the number of model parameters,improve the training speed and more effectively extract rich multi-scale features;Finally,the decoder refinement module is designed to restore the feature map.Experimental results on Cityscapes and PSCAL VOC 2012 datasets show that the model achieved a faster segmentation rate and higher segmentation accuracy based on the lower number of parameters,and MIo U improved by 2.31% and 3.15% compared with the original model.