Research On Multi-period Infrared Image Generation Based On SERED-GAN

Infrared imaging technology has the advantages of unlimited working hours and relatively high resolution.It has been widely used in various fields such as navigation.The acquisition of infrared images requires huge financial and time costs,and infrared image at different time periods has many differences,making it difficult to acquire infrared images at all time periods.How to quickly and efficiently generate high-fidelity multi-period infrared images has become the focus of current research.The image-to-image translation model based on Generative Adversarial Network has greatly surpassed the traditional method in image generation tasks in recent years.This article applies this idea to infrared image generation tasks in modeling the conversion relationship between visual images and infrared images,thus infrared simulation images can be obtained from the easily obtainable visible light images through image-to-image translation.Early GAN-based image-to-image translation networks mostly required pixel-level matched image pairs for training,and hard to generate multi-period infrared images.To address this problem,this article introduces the unsupervised learning model Cycle GAN to perform image-to-image translation.This method relieves the pixel-level matching constraints on training data by imposing cycle consistency on the network,making the translation model retain the scene information.To address the problem of multi-period infrared image generation,this article uses Star GAN network.The labels of the target time period are input of the generator,while the discriminatory classifies the generated images into different time period.In this way,a unified model is built to the perform image-toimage translation between visible images and multi-period infrared images,lessening the tedious steps of training multiple translation models.Due to the large gap between visual and multi-period infrared images,it is difficult to accurately model the translation relationship with an end-to-end generation network such as Star GAN.To address this problem,this article proposes a multi-period infrared image generation method based on Self-supervised Representation Disentangling,called SEREDGAN.The generator in SERED-GAN is divided into coding stage and decoding stage,and two networks are established separately for these two stages,which enhances the expression ability of the model and reduces the difficulty of end-to-end training.In the encoding stage,self-supervised learning is used to disentangle the image's content features and style features,and then recombine the visual image's content features and infrared image's style features,and finally decode the reconstructed image features to restore the target image.Finally,extensive experiments are conducted on the public vehicle front view infrared image dataset KAIST to compare the SERED-GAN model with other image-to-image translation methods.The experimental results show that the SERED-GAN proposed in this article can achieve or even exceed the supervised learning model in the quality of the generated image.It can model the characteristics of multi-period infrared image to generate multi-period infrared images with high fidelity.Finally,experiments are conducted to test the model's heterogeneous data generalization performance.The results point out the shortcomings of the current research and the direction of future research.