Research On Image Semantic Segmentation Algorithm By Boundary Assistant Network

Image semantic segmentation has always been a fundamental and challenging task in the computer vision,and it plays an important role in many practical applications,such as robotics,and medical segmentation.In recent years,the application of Convolutional Neural Networks(CNNs)in the field of semantic segmentation has made significant progress.For example,methods based on fully convolutional network(FCN)and based on encoder-decoder architecture(EDA).Although these methods have made some progress in segmentation tasks,they all suffer from rough segmentation of object shapes and boundaries.However,for the semantic segmentation task,it is not only necessary to accurately identify the category of each pixel in the image,but also to accurately locate its position,which requires the model to accurately identify the contour and shape of each category in order to obtain a continuous semantic segmentation prediction result..Furthermore,although the accuracy of semantic segmentation is improved by refining object boundaries,this progress is achieved at the cost of substantial computational resources.Based on the above problems,this thesis mainly conducts the following research:(1)In order to effectively use boundary information to assist in improving the performance of semantic segmentation,this thesis proposes a two-branch symmetric network(BASNet).It mainly consists of four parts: Based on pre-trained residual network(Res Net-50/101)plus ASPP(Atrous Spatial Pyramid Pooling)as backbone,Semantic segmentation branch,Boundary detection branch and Aggregation module.Specially,Boundary detection branch processes boundary-related information using a series of SABs(Spatial attention blocks).Besides,a set of GABs(Global attention blocks)is used in the semantic segmentation branch to further capture more accurate position information and semantic information.Finally,BANet introduces an AM to effectively fuse the output features of these two branches,thereby improving the final semantic segmentation performance.Comprehensive experimental results show that BASNet not only predicts object boundaries more accurately,but also improves the performance of semantic segmentation.(2)Based on the research and analysis of the strengths and weaknesses of the previous work,this thesis designs a boundary-assistant encoder-decoder network for semantic segmentation(BANet),which is a pre-trained residual network(Dilated-Res Net-50/101)on the encoder,the decoder consists of boundary flow branch(BFB)and semantic flow branch(SFB).The semantic flow branch uses a series of bilateral global attention modules to make full use of hierarchical features at different stages,while capturing long-distance dependencies between pixels from the horizontal and vertical directions,respectively.By adopting a series of lightweight spatial attention modules,the boundary flow branch can not only effectively extract boundary information,but also its module parameters and computational complexity are lighter than previous work.Comprehensive experimental comparison,BANet achieves m Io U of 83.8%,55.3% and 49.4% on Cityscapes,PASCAL-Context and ADE20 K,respectively.Furthermore,due to the lightweight design of the lightweight spatial attention module,BANet requires lower GFLOPs and smaller model size relative to some state-of-the-art.(3)Currently,in order to be able to deploy CNNs in edge devices,not only the models need to have low-latency real-time processing capabilities,but also the image semantic segmentation algorithms applied on them have strong performance.Therefore,on the basis of the previous work,this thesis proposes a real-time semantic segmentation network based on boundary assistance(Lite-BANet).By introducing an extremely efficient cross-stage non-bottleneck residual module(EECSNB)to enhance the feature expression ability of each scale receptive field at different stages,the module uses channel split operation and factorized convolution to reduce the amount of parameters of the model In addition.By partially optimizing the two branches in BANet,the boundary information and semantic information at different scales are further learned,so as to achieve a balance between model accuracy and speed.The experimental results show that the network achieves 76.4% m Io U and 101.2 FPS on the Cityscapes dataset,achieving a good balance between accuracy and speed.