Research On Image Segmentation Model Based On Position And Channel Attention Mechanism

With the rapid development of neural networks,semantic segmentation based on neural networks has achieved a huge breakthrough in effect compared with traditional methods.Semantic segmentation technology can be applied to a variety of scenarios,such as unmanned driving,medical image analysis,virtual reality,and human-computer interaction.Researchers continue to improve the network structure to improve the performance of neural networks in semantic segmentation.It is an effective way to add attention mechanism to the network structure.Existing self-attention networks obtain global feature sharing through a large number of matrix multiplications,so they have a large amount of parameters.This paper aims to improve the existing attention mechanism to improve the segmentation effect of the network.Mainly did the following work:By improving the self-attention module,this paper proposes a new type of Position Feature Sharing module(PFS)to characterize the contextual relevance of each pixel location to all global pixel locations.Before the original matrix is multiplied to obtain the weight of the relationship between the positions,a global average pooling is performed,which reduces the amount of parameters without affecting the performance.For the channel attention module,this paper proposes a Channel Feature Sharing module(CFS),which abandons the original matrix multiplication,first minimizes the impact of location information through global average pooling,and then obtains it through a full connection.The context relationship between each channel achieves better performance on the basis of reducing the amount of parameters.Based on the above two modules,the semantic segmentation model of this paper is obtained: Dual Feature Sharing Network(DFSNet).In this paper,based on DFSNet,through multi-scale design,the channel feature sharing module is added to the convolution branch of different scales to capture the channel context relationship at different scales,and a Multi-scale Feature Sharing network(MFSNet)is designed,which is carried out with DFSNet.Compared with experiments,performance has been improved.This article is implemented on the City Scapes,PASCAL VOC 2012,and COCO stuff data sets,and compares the improved network in this article with the current more classic network model,and obtains a higher average intersection ratio(m Io U).This paper also explores the serial and parallel connection of the channel feature sharing module and the location feature sharing module,as well as the sequence of the serial connection.It is found that the Channel Feature Sharing module is more effective in series connection.Experiments show that the model has good convergence and robustness,and has a better segmentation effect.