Research On Image Semantic Segmentation Based On Attention Mechanism And Feature Fusion

Semantic segmentation is a major task in computer vision,whose basic task is to label each pixel according to the real objects that exist in the image.Image semantic segmentation technology can be applied to many fields such as medical image analysis,geographic information systems,unmanned driving,and so on.Taking unmanned driving as an example,intelligent control systems need to rely on computer vision to continuously perceive changes in spatial information,plan paths and form driving instructions in real time.Semantic segmentation is the basic technology and method for identifying and sensing spatial information,and the safety of unmanned driving puts forward higher requirements for the accuracy of semantic segmentation.Currently,semantic segmentation methods based on deep learning often have problems such as low segmentation accuracy for small scale objects and insufficient continuity for local boundaries.Therefore,this thesis proposes an image semantic segmentation method based on attention mechanism and feature fusion.The main research purpose is to improve the segmentation accuracy for small scale objects and target boundaries while ensuring the overall segmentation accuracy,Meet the requirements of accurately segmenting objects in complex backgrounds.The main innovations and research results of this article are as follows:(1)An image semantic segmentation model based on hybrid concatenation and feature fusion is proposed.The overall architecture of the network model adopts an encoder decoder structure,and the hole space pyramid pooling module is improved in the encoder.The hybrid concatenation and multi-core pooling methods are used to extract deeper semantic information,better integrating global multiscale context information,and improving the overall segmentation accuracy of the target.(2)A cross stage fusion method is designed to divide the backbone network of the encoder stage in the network model and the improved hole space pyramid pooling module into three stages:shallow,middle,and deep.By using a jump connection method,the feature maps of each stage are fused,enabling the network model to fully utilize the different semantic information of the shallow and deep layers,and enhancing the dependency between features,Thereby improving the segmentation accuracy of small scale objects.(3)An image semantic segmentation model based on cross stage and refined attention mechanisms is proposed.Attention mechanisms are introduced into a hybrid cascade and feature fusion image semantic segmentation network model,and improvements are made on the convolutional attention mechanism(CBAM).This method can help the network model focus on meaningful features.Adding self attention to channel attention enhances the connection between feature maps,and using one-dimensional convolution in the spatial attention mechanism to increase the spatial receptive field,enriching the information of feature maps and enhancing the representation ability of network models to extract features.The method in this article has been trained and tested on the public dataset PASCAL VOC2012 and SUIM.MIoU has reached 86.68%on PASCAL VOC2012 and 61.55%on SUIM.Experiments have shown that the overall accuracy of the method in this thesis is higher than many existing methods,which proves the effectiveness of the method in this thesis.