Research On Image Semantic Segmentation Based On Fully Convolutional Neural Network

Image semantic segmentation is a very important research field in computer v ision,and it plays an extremely important role in the analysis and understanding of image contents.Image semantic segmentation can classify each pixel according to different semantic meanings in an image,so that pixels belonging to the same type of object are divided into the same group.In recent years,with the emerge nce and development of fully convolutional neural n etworks,image semantic segmentation technology has made great progress.However,the existing image semantic segme ntation methods based on fully convolutional neural networks still have major pro blems such as difficulty in segmenting multi-scale objects correctly,loss of a large amount of spatial information,and lack of context information.Aiming at the problems existing in semantic segmentation and its research status,this paper designs more effective image semantic segmentation models and algorithms to further improve segmentation accuracy.The research work of this paper includes the following aspects:(1)In order to solve the problems of low segmentation accuracy due to lack of context information,a segmentation method based on convolution dense pyramid pooling is proposed in this paper.This paper combines atrous spatial pyramid pooling module and spatial CNN module,and embeds them into the network in a dense co nnection way,so that it can extract more effective multi-scale context information,so as to better capture objects of different scales in an image.(2)In order to solve the problems of losing a lot of spatial information and di fficult to correctly segment multi-scale objects in the segmentation methods,this paper proposes a segmentation method based on residual dilated attention to solve these two problems at the same time.The residual dilated attention model mainly uses series and parallel ways to deploy dilated convolutions,and then design reasonable sampling rates for these dilated convolutions,and finally connect them in the form of residuals.By retaining more spatial information and aggregating multi-scale feature information,the model can effectively segment objects of di fferent scales in an image.(3)In this paper,a lot of experiments have been carried out on the two proposed image semantic segmentation methods.The experimental results verify the effectiv eness of the proposed segmentation methods,and show that the prop osed segmentation methods have competitive performance in image semantic segmentation.