基于像素语义信息的单图像视图生成深度预测方法  

SINGLE-IMAGE-BASED VIEW GENERATION AND DEPTH PREDICTION METHOD BASED ON PIXEL SEMANTIC INFORMATION

作　　者：孔维罡 郭乃网 周向东[1] Kong Weigang;Guo Naiwang;Zhou Xiangdong(School of Computer Science,Fudan University,Shanghai 200433,China;State Grid Shanghai Municipal Electric Power Company,Shanghai 200437,China)

机构地区：[1]复旦大学计算机科学技术学院,上海200433 [2]国网上海市电力公司电力科学研究院,上海200437

出　　处：《计算机应用与软件》2023年第4期192-198,235,共8页Computer Applications and Software

基　　金：国家电网科技项目(520940180029)。

摘　　要：单幅图像的深度预测具有重要的应用前景。为解决现有视图生成方法中图像扭曲的问题,创新性引入目标检测思想改进了视图生成网络Deep3D,提出基于像素语义信息的单图像视图生成模型。把视图生成网络得到视差概率分布和目标检测模型提取到中心点概率分布加权结合到一起。把模型得到的视差图与输入图结合来产生右图,最后利用左右图计算得到深度图。实验结果显示,该方法有效提升了生成右图和计算得到深度图的精度。The single image depth prediction has great potential in many practical applications.In order to deal with the image distortion problem of the existing methods,we improved the view generation model Deep3D through innovative introduction of target detection idea.exploring pixel semantic distribution.A single-image-based view generation model based on pixel semantic information was proposed.The disparity probability distribution obtained by the view generation network and the center point semantic distribution from the target detection model were weighted together.as the probability model for estimating the new disparity map.With the input image and obtained disparity map,a right view can be obtained.The image depth map was generated by using both the left and right views.The experimental results show that this method generates more accurate right view and improves the depth estimation significantly.

关 键 词：视图生成 深度预测 目标检测 语义信息 概率模型 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]