Research On Semantic Segmentation Based On Attention And Information Fusion

Semantic segmentation is an important direction of computer vision research.Different from the rough classification of traditional image recognition,semantic segmentation can make pixel-level intensive prediction of images,so that it can be more adapted to the needs of realworld applications.In recent years,semantic segmentation technology has developed rapidly,and new methods have emerged one after another.Among them,the use of attention thinking and fusion thinking has gradually become the mainstream direction of the development of semantic segmentation.This article starts with the thought of attention and the idea of fusion,and makes corresponding improvements and innovations to solve the problems of imprecise segmentation of specific objects,imprecise boundary segmentation,large memory usage and slow operation speed in current semantic segmentation.The research content of this article is as follows:(1)Aiming at the problem of insufficient feature extraction capabilities of traditional backbone feature extraction networks,this paper studies the impact of different backbone feature extraction networks on the accuracy of semantic segmentation.Finally,the superior performance of Res2Net,a backbone feature extraction network with multi-scale fusion and residual ideas as the core,is verified.(2)Aiming at the problem of difficult to segment difficult samples,this paper improves the loss function,and uses the improved loss function to effectively improve the accuracy of segmentation.(3)This article conducts related research on attention thinking.In the study of attention mechanism,this paper improves the traditional spatial attention and channel attention modules,and designs a new attention module for horizontal and vertical spatial information.In the study of the self-attention mechanism,this paper improves the computational complexity of spatial self-attention and the large memory usage,effectively reducing the complexity of the spatial self-attention module by one level.(4)This article conducts related research on the idea of information fusion.Among them,in the research of multi-dimensional feature fusion,this paper adds stream alignment module to perform feature alignment before feature fusion,which greatly improves the effect of multidimensional feature fusion.In the study of multi-scale feature fusion,based on the current feature map size,the current existing multi-scale feature fusion method is adjusted with hyperparameters,and combined with the attention idea,a multi-scale based on self-attention mechanism is proposed.Feature fusion module.In addition,using category context information to correct network error classification,this paper proposes a new multi-branch information fusion method.(5)In this paper,the current semantic segmentation network has a large computing resource occupancy and cannot be used normally due to its slow computing speed.This paper conducts a research on a lightweight semantic segmentation network.This paper builds a twobranch lightweight semantic segmentation network based on the idea of information fusion and attention,and uses the Ghost module to further reduce its parameter;this article uses the optimal model in the previous article as the teacher model for knowledge distillation training,thus Transfer the knowledge in the teacher network,use the KL divergence as the secondary loss,and use the dynamic temperature adjustment strategy in the knowledge distillation training.