Research On Image Semantic Segmentation Method Based On Attention Mechanism

In the information age of the rapid development of high technology in the 21 st century,Image semantic segmentation,one of the technologies in the field of computer vision,plays an extremely important role.Image semantic segmentation can be used in unmanned driving,medical images and other visual fields.However,the Image Semantic Segmentation still exist many problems need to be solved,such as rough semantic feature fusion,semantic information is inconsistent,the current image semantic segmentation method are difficult to consistent semantic prediction for image target,in order to improve the image semantic segmentation method for overall generalization performance of image objects and the semantic consistency,In this paper,Res Net is used as the backbone of network feature extraction to carry out the research of image semantic segmentation based on attention mechanism.(1)Aiming at characteristics of fusion in image semantic segmentation is rough,simple high-level information and the underlying message in addition,there will be a rough feature fusion,such as edges and detail information is relatively vague,and even the boundaries between objects disappear.For this problem,Res Net is used as the backbone network for feature extraction,the use of the avg pooling and max pooling and nonlinear activation function Sigmoid operation,design the Pooling Coordinates Attention Network,on the transverse and longitudinal directions,while considering the channel attention information,attention information embedded in the corresponding position,under the effect of max pooling and avg pooling,The maximized feature extraction of image semantic information is completed.At the same time,the channel attention information and the corresponding location attention information are considered,which solves the problem of missing object edge and obscure detail information.Experiments are carried out on PASCAL VOC 2012 and Cityscapes datasets,and the experimental performance reach the best 74.4% MIo U and 74.6% MIo U.At the same time,a more intuitive visualization result is used to verify that the Pooling Coordinates Attention Network has obvious effect on solving the problem of missing edges and obscure details.(2)Aiming at semantic prediction in image semantic segmentation inconsistency problem,most of the semantic network segmentation and the correlation between pixels,resulting in image semantic segmentation task,predicted results for segmenting the image objects lost the consistency of semantic information.In order to solve the weak correlation between pixels to this problem,Res Net is used as the backbone network for feature extraction,this paper first use the avg pooling and max pooling and residuals to extract the contour information of network structure,the details information and shallow attention information,in order to enhance the correlation between the image pixels,and then on the basis of this,further combined with the matrix multiplication,the Cascade Channel Attention Network is designed to enhance the correlation between image feature semantic information by identifying the interdependence between salient channels and modeling channels.Experiments are performed on PASCAL VOC 2012,Cityscapes,MS COCO 2014 and other mainstream datasets.Among them,PASCAL VOC 2012 and Cityscapes data sets respectively reach the best 77.1% MIo U and 73.6% MIo U.Finally,visualization results are used for intuitive analysis.The designed network has a significant effect on enhancing the correlation between the semantic information of image features.On the basis of the above,the experimental results are compared with some advanced image semantic segmentation networks.Through the comparison of experimental data and the analysis of visualization results,the proposed image semantic segmentation method has better semantic segmentation performance.