Research On Image Semantic Segmentation Method Based On Adaptive Non-local Attention Network

Image semantic segmentation is an important technology in computer vision task with great application value and research prospects.By assigning the same label to the pixels corresponding to target class in given image,the application device to automatically recognize each target class in a given image.Due to the design of the convolutional neural network structure,the receiving field is limited to the local area of the convolution,which leads to the problem of insufficient understanding of context information.Therefore,a reasonable context information aggregation method is the basis of establishing an efficient scene image semantic segmentation model.In this paper,an image semantic segmentation method based on an adaptive non-local attention network is studied and implemented by integrating various theories and methods such as attention mechanism,multi-scale features,secondary information dissemination,etc.The main research results of this paper are as follows:(1)Due to the loss of detail information caused by continuous convolution layer or pooling layer,the current semantic segmentation methods generally use one-way fusion,which is limited to the one-way fusion of deep features to help restore the local feature map.The importance of location information in local features is ignored.In order to enrich the diversity of feature information,a Bi-directional Feature Fusion Block(BFFB)based on the fusion of global features and local features is proposed in this paper.This module helps to restore the semantic information of local features by learning global features through deep network.At the same time,it learns local features through shallow network,helps to restore the location information of global features,resolves the problem of missing local features,and makes the network understand global context information more fully.(2)The current image semantic segmentation methods use the homogeneous context dependence of all image regions in a non-adaptive manner.The differences between the local representation and context dependence of different categories and the problem of blurring the edges of the segmentation results are ignored.This paper uses remote dependence to enrich non-local context information,and proposes Attention Information Gathering Block(AG)and Attention Information Distribution Block(AD).AG gathers the characteristic information of other relevant points to seek commonality,and AD distributes current location information to other relevant points to distinguish personalities.The problem of edge blurring in segmentation results is solved by generating attention maps as weights to help feature adaptive learning.(3)This paper proposes a general semantic segmentation architecture-Adaptive Non-local Attention Network(ANANet).The network is composed of an improved backbone network based on Res Net and an ANA information aggregation module,which can adapt to input images of different resolutions for end-to-end training.Between them,the ANA information aggregation module is integrated by BFFB,AG,and AD.ANANet is tested on Pascal VOC 2012 data set and Cityscapes data set,and compared with FCN,Deep Lab and other classic networks in the same environment.The results show that ANANet has achieved excellent results in the two data set experiments,and is superior than the previous network.