Research On Digital Camouflage Generation Based On Generative Adversarial Network

With the development of military technology,digital camouflage as an important protection method is widely used in the confrontation of military reconnaissance technology.The existing digital camouflage generation methods usually need a series of steps to complete the generation of digital camouflage,such as the main color extraction of the background image,feature selection,digital patch generation and color filling.Although the generated digital camouflage is more delicate,it is unable to generate digital camouflage end-to-end from the background.And the generated digital camouflage is time-sensitive.When the season changes,you need to follow the above steps to regenerate the digital camouflage.At the same time,the main goal of the traditional digital camouflage generation method is to destroy the edge information of the background image to achieve camouflage for the human eye.There is no guarantee that the generated digital camouflage can also have a higher camouflage when facing the neural network image classification model.Success rate.In response to the abovementioned problems,this article combines the generation of confrontation networks to carry out the following three researches on the generation of digital camouflage:Aiming at the problem that traditional digital camouflage generation methods cannot perform end-to-end digital camouflage generation,a digital camouflage generation method based on cyclic consistency against the network is proposed.The image features are extracted through densenet,elu and selu are added as activation functions to the generator and discriminator to improve the quality of the generated digital camouflage.The image structure retention loss is added to ensures that the generated digital camouflage maintains the structure information of the original background image,finally realize end-to-end digital camouflage generation.Regarding the problem that the traditional digital camouflage generation method cannot generate digital camouflage with different seasonal characteristics under the same background conditions,a digital camouflage generation method based on disentangled representation is proposed.This method embeds different images onto two spaces: a domain-invariant content space capturing shared information across domains and a domain-specific style space.The content features from the background image and digital camouflage are aligned,and the aligned background image content features are combined with the digital camouflage style features to generate digital camouflage patterns with different seasonal styles under the same background.Regarding the problem that the generated digital camouflage cannot effectively counter the image classification model,a kind of adversarial sample category loss for a specific image classification model is proposed.Combine the camouflage target with the digital camouflage,use the loss function to guide the classification model to output the specified error classification,and realize the adversarial attack against the specific image classification model.The experimental results show that the image after adversarial camouflage achieves a better camouflage effect when facing a specific classification model.