Research On Infrared Image Generation And Super Resolution Base On Generative Adversarial Network

Since the 20th century,infrared imaging technology has been widely used and developed.It has been successfully used in many fields,such as military field,to observe in the dark.The corresponding infrared image target detection technology has also been the corresponding development,from the traditional edge feature extraction,threshold segmentation and other methods,to now based on deep learning and neural network method is widely used.However,infrared imaging technology is still faced with the problems of low imaging quality and limited observable features.In addition,in military applications,the target often presents small and fuzzy features in the image,so it is difficult to detect the target directly.Since generative adversarial networks was proposed in 2014,this direction has gradually become a hot spot of artificial intelligence research,and has been applied in various fields,such as image generation,style migration,image reconstruction and so on.As a pilot experiment of infrared and visible image fusion target detection,this paper studies the expansion and super-resolution reconstruction of the existing infrared image data set of small targets in the air based on the generative adversarial network(GAN).The main contents and research points of this paper are as follows:(1)firstly,the principle of the main algorithms used in the experiment is introduced,including the generative adversarial network,attention mechanism,target detection and so on.(2)Based on the self-attention GAN,the original image and the cut out foreground image block are learned by different discriminators to learn the features of small target foreground and complex background at the same time,and then the image is fused to generate more realistic data with the target object,and the data is expanded on the basis of the existing infrared data set.At the same time,a target extractor based on YOLO framework is pretrained to detect small targets in the image.The final results in is,FID,MMD and other indicators have more than 2% improvement,the detection result of YOLO at 0.5 threshold is more than twice,which shows that the expanded image has good detection results.(3)In the framework of SRGAN,attention mechanism is introduced to reconstruct the infrared image with lower resolution,so as to generate clearer and higher quality image,which is conducive to the realization of target detection task.The experimental results show that the super-resolution image not only improves PSNR and SSIM by more than 0.3%,but also improves the image quality in subjective perception.