基于深度流场特征和语义约束的改进SegFormer语义分割算法  

Improved SegFormer Semantic Segmentation Algorithm Based on Deep Flow Field Feature and Semantic Constraint

作　　者：高延海 GAO Yanhai(Qingdao University of Technology,Qingdao 266520,China)

机构地区：[1]青岛理工大学,山东青岛266520

出　　处：《现代信息科技》2024年第18期71-74,共4页Modern Information Technology

摘　　要：对于SegFormer网络中存在的多尺度信息无法有效利用,以及预测结果边界轮廓不清晰的问题,提出一种基于深度流场特征和语义约束的改进语义分割网络架构。首先,在解码器部分加入深度流场模块用于加强深度特征的一致性;然后为保持原有网络的轻量化,加入边界和前后景辅助任务构成语义约束模块,提高网络对边界和整体轮廓的提取能力;最后,在语义约束模块中加入边界引导模块,加快辅助任务收敛速度。通过增加了0.1 M的参数量,提高了网络的分割精度。For the problems that the multi-scale information existing in the SegFormer network cannot be effectively utilized and the boundary contour of the prediction result is unclear,an improved semantic segmentation network architecture based on deep flow field feature and semantic constraint is proposed.Firstly,a deep flow field module is added to the decoder part to enhance the consistency of depth feature.Then in order to keep the lightweight of the original network,the boundary and foreground and background auxiliary task are added to form a semantic constraint module to improve the ability of network to extract boundary and overall contour.Finally,the boundary-guided module is added to the semantic constraint module to speed up the convergence of auxiliary task.By increasing the number of parameters by 0.1 M,the segmentation accuracy of the network is improved.

关 键 词：SegFormer 语义分割 轻量化 深度流场 辅助任务 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]