Image Semantic Segmentation Network Based On Deep Learning And Boundary Optimization

As an important task in the field of computer vision,image semantic segmentation plays an important role in medical image segmentation,automatic driving,industrial inspection and other fields.In recent years,deep learning has been widely used in image semantic segmentation.For example,the convolutional neural network with encoderdecoder structure can achieve high-precision image segmentation.However,a major defect of the existing methods is that they cannot achieve a good balance between segmentation accuracy and model size.In addition,some algorithms lack effective use of boundary information and cannot better identify the boundaries of objects.To this end,this article will start from image semantic segmentation,focusing on application of lightweight technology in deep learning and improvement of object shape by boundary optimization technology.The specific research contents are as follows:(1)Lightweight Deep Labv3+ semantic segmentation network.In order to make Deep Labv3+ network better applied to mobile devices in terms of model size and inference speed,an improved lightweight Deep Labv3+ semantic segmentation network is proposed.Proposed network first replaces the backbone network with Mobile Netv2,then improves original atrous spatial pyramid pooling module and introduces ECA module to weight the channels after improving pyramid pooling module.Finally low-level information is supplemented through multi-branch feature fusion to improve segmentation performance.Experimental results show that on PASCAL VOC 2012 and Cityscapes datasets,improved Deep Labv3+ network not only greatly reduces the number of parameters of the model,but also improves final segmentation performance,with m Io U increased to 73.31% and 75.42%,respectively.(2)Attention-based boundary refinement semantic segmentation network.In order to further improve the segmentation accuracy,it is necessary to use boundary information to help the network improve the results of semantic segmentation,a semantic segmentation network combining boundary branches is proposed on the basis of lightweight Deep Labv3+ semantic segmentation network.The network mainly consists of four parts: Lightweight Deep Labv3+ semantic segmentation network as backbone network,semantic branch,boundary refinement branch and attention-based fusion module.Among them,the backbone network and semantic branch are the same as previous work.Boundary refinement branch fuses output features from different stages of backbone network.Besides,network guides boundary refinement branch to learn boundary-related information with an additional boundary loss function.Finally,the output features of the two branches are fused by fusion module to improve final segmentation performance.The Experimental results show that the proposed network improves the boundary of segmented objects by adding boundary refinement branch,and improves semantic segmentation performance of the network.On PASCAL VOC2012 dataset,m Io U reaches 73.88%.(3)A lightweight boundary-optimized semantic segmentation network based on dual branches.Considering that information of boundary is relatively simple,it can be represented by fewer channels.Therefore,in order to further lightweight the semantic segmentation network,this paper proposes a lightweight boundary-optimized semantic segmentation network based on dual branches.Proposed lightweight processing network is based on attention-based boundary refinement semantic segmentation network.A new boundary branch module combined with attention mechanism is proposed to optimize boundary.At the same time,lightweight processing is performed on semantic branch and guides the learning of network through attention mechanism.Finally,the outputs of the boundary branch and the semantic branch are fused by the proposed attention-based fusion module.Experimental results show that lightweight dual-branch semantic segmentation network not only greatly reduces the amount of parameters and calculations of the model,but also improves the segmentation result to a certain extent: m Io U reached 73.40% on the PASCAL VOC 2012 dataset.