Research On Infrared Image Generation Method Based On Generative Adversarial Network

Infrared imaging is widely used in civil and military fields,which has strong antijamming ability and can obtain clear images from a distance and in night scenes.In the process of developing infrared-related equipment,a large number of infrared images under various conditions are needed as verification data.However,it takes a lot of manpower and material resources to acquire real-time infrared images in the field.At the same time,it is difficult to obtain infrared images at all times.In order to solve the problem of insufficient infrared image datasets,this paper introduces the generative adversarial networks into infrared image generation task,and studies infrared image generation method based on visible image and infrared image period extension technology.The main contributions of this thesis are as follows:(1)Aiming at performance degradation of generating infrared images caused by haze on visible images,this paper proposes a dehazing algorithm based on image fogging degree.Considering that the transmission map reflects the fogging degree of hazy images,we build a transmission map prediction module parallel to the dehazing module.The threshold adaptive learning module is designed to calculate the segmentation threshold.According to the segmentation threshold,we can get mask maps of different fogging degree,which can guide the separation defogging convolution module to recover clear images.Experiments show that our algorithm achieves good defogging effect on both public datasets and natural images.(2)In order to generate infrared images of corresponding scenes based on easily available visible images,this paper tries to apply the Pix2 pix network to the paired visible-infrared image datasets.To solve the problem of missing detail information of infrared image generated by Pix2 pix network,this paper proposes the multi-receptive field feature fusion Pix2 pix network.A multi-receptive field features extractor based on the Unet ++ structure is constructed and a multi-receptive field feature fusion mechanism is proposed.Experiments show that the multi-receptive field feature fusion Pix2 pix network achieves finer infrared texture generation.(3)Aiming at the lack of paired multi-period infrared image datasets,this paper firstly proposes an infrared image period extension algorithm based on Star GAN network.In order to solve the problem that Star GAN network gets different results in expanding the same material image region,this paper proposes the semantic constrained Star GAN network.On the basis of the Star GAN network,a semantic segmentation branch is added.We inject the scene semantic features into net and design the semantic consistent loss.Aiming at the problem that the result of the expansion of some material image regions by Star GAN network is wrong,this paper improves the period-coding method of the semantic constrained Star GAN network.We propose a period-coding method based on semantic segmentation maps and infrared radiation time-varying curves.Experiments show that the semantic constrained Star GAN network is more suitable for the task of infrared image period extension.