基于改进pix2pix的红外图像转换技术
Infrared image conversion technology based on improved pix2pix
叶明亮 1史春景 1郝永平 2李大伟1
作者信息
- 1. 沈阳理工大学机械工程学院,辽宁沈阳 110159
- 2. 沈阳理工大学装备工程学院,辽宁沈阳 110159
- 折叠
摘要
针对不同波段图像获取代价不同的问题,提出一种基于pix2pix的图像转换方法并进行改进.主要针对生成器和鉴别器两方面进行改进.生成器方面,使用残差结构的生成器替换原来的U-Net生成器以缓解梯度消失问题;引入可变形卷积,提高目标边缘和小目标的生成效果;引入BAM注意力机制,提高了算法对图像中主要目标的特征提取能力以提升生成图像的效果.鉴别器方面:改变PatchGAN中卷积层的层数(原PatchGAN为3层卷积),设置对照实验找到转换效果最好的卷积层数.以可见光图像和红外图像之间的转换为例进行实验.实验结果表明,改进后的算法在生成图像上的均方根误差(MSE)下降了 31.4％、结构相似性(SSIM)提高了 11.2％,可以更好的实现红外图像和可见光图像之间的转换.
Abstract
In order to solve the problem of different cost of image acquisition in different light segments,an image con-version method based on pix2pix was proposed.It mainly focuses on the generator and discriminator.In terms of gener-ators,the residual structures generator was used instead of the original U-Net generator to alleviate the gradient vanis-hing problem.Deformable convolution is introduced to improve the generation effect of target edges and small tar-gets.The BAM attention mechanism is introduced to improve the feature extraction ability of the algorithm for the main target in the image to improve the image generation effect.In terms of discriminators:change the number of convolu-tional layers in PatchGAN(the original PatchGAN is 3-layer convolution),and set up a control experiment to find the convolutional layer with the best conversion effect.Some KAIST datasets are selected for training and testing.The ex-perimental results show that the Root Mean Square Error(MSE)of the improved algorithm is reduced by 31.4％and the Structural Similarity(SSIM)is increased by 11.2％,which can better realize the conversion between infrared and visible images.
关键词
生成对抗网络/pix2pix/图像转换/残差结构Key words
generative adversarial network/pix2pix/image conversion/residual structures引用本文复制引用
出版年
2024
NETL
NSTL
万方数据