基于生成对抗网络的船舶可见光图像到红外图像的转换  

Transform from visible ship image based on generative adversarial network to infrared ship image

作　　者：王志坚 陈春雨[1] WANG Zhijian;CHEN Chunyu(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001

出　　处：《应用科技》2024年第5期243-248,共6页Applied Science and Technology

基　　金：中央高校基本科研业务费项目(3072020CFT0803)。

摘　　要：为了解决红外制导研究中舰船图像样本数量不足的问题,提出一种面向舰船图像的改进的生成对抗网络(generative adversarial network,GAN),能够生成高质量的红外图像。首先转换可见光图像颜色空间以更好地捕捉夜间低亮度下图像的轮廓信息,然后引入残差块生成网络降低低像素的可见光图像对生成的红外图像的影响并加深网络层数以更好地学习深层映射关系,最后引入更平滑的损失函数加快收敛速度,提高生成红外图像目标边缘清晰程度。在制作的无人机拍摄的红外可见光配对的数据集进行测试,改进后的方法平均生成图像峰值信噪比(peak signal to noiseratio,PSNR)提升20.3%,结构相似性度量(structural similarity,SSIM)提升30.4%。结果表明改进的网络可以生成质量更高的红外仿真图像,用于目标检测等任务有更好的效果。To address the issue of insufficient image samples of ship images for infrared guidance research,an improved generative adversarial network(GAN)is proposed specifically for ship images.This network is capable of generating high-quality infrared images.Firstly,the color space of visible light images is transformed to better capture image contour information under low-light conditions;Then residual block generation network is introduced to reduce the impact of low-pixel visible light images on the generated infrared images,deepen the network layers for better learning deep mapping relationships;Finally,a smoother loss function is introduced to accelerate convergence speed and enhance the edge clarity of the generated infrared image.The test on a dataset of infrared-visible light pairs captured by unmanned aerial vehicles showed that the improved method had an average peak signal-to-noise ratio(PSNR)improvement of 20.3%and structural similarity(SSIM)improvement of 30.4%.Experimental results indicate that the enhanced network can generate higher-quality simulated infrared images,leading to better performance in such tasks as target detection.

关 键 词：域转换 对抗生成网络 红外图像生成 颜色空间 残差块 平滑损失函数 峰值信噪比 结构相似度 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]