一种基于生成对抗网络的火焰图像场景迁移新方法  

A new scene migration method for fire images based on generative confrontation network

作　　者：王腾[1] 杨植凯 卜乐平[1] 侯新国[1] 欧阳继能 WANG Teng;YANG Zhi-kai;BU Le-ping;HOU Xin-guo;OUYANG Ji-neng(College of Electrical Engineering, Naval Univ. of Engineering, Wuhan 430033, China;Unit No. 92823, Sanya 572000, China)

机构地区：[1]海军工程大学电气工程学院,武汉430033 [2]92823部队,海南三亚572000

出　　处：《海军工程大学学报》2022年第2期27-32,56,共7页Journal of Naval University of Engineering

基　　金：国家自然科学基金资助项目(41971416,41771487)。

摘　　要：因为在一些特殊场景(如舰船舱室)下禁止进行点火实验,导致该场景下的火灾视频数据不足,严重制约了视频火灾探测模型在该场景下的有效应用。为解决该问题,提出了一种基于生成对抗网络的火焰图像场景迁移方法。该方法克服了现有方法将图像信息解缠绕至隐空间的缺陷,能在特定的无火场景视频中添加火焰并完成风格信息的融合,保持了合成图像高质量的背景细节。实验结果表明:该方法能够生成不同形态的火焰图像,并使迁移后的火焰场景图像具有较高的视觉真实性。In some special scenes such as ship cabins,the fire video sampled in this scene is insufficient due to the prohibition of ignition,which severely restricts the effective application of the video fire detection model in these scenes.In order to solve the problem,a flame image scene migration method based on a generative confrontation network was proposed,which overcomed the defect of the existing method of unwinding the image information into the hidden space,added flames to a given scene video and completed the fusion of style information while maintaining high quality background details.Experimental results show that the method ensures the diversity of flames in the scene and produces perceptually realistic results.

关 键 词：生成对抗网络 火焰图像生成 场景迁移 注意力机制 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]