Research On Intelligent Camouflage Technology For Remote Sensing Image Target

In today’s society,the use of deep learning for target recognition and big data analysis of remote sensing images has gradually become a trend.The traditional camouflage design method for destroying edge information is a general-purpose network design algorithm for a specific background.It cannot guarantee the effect on the designated detection network and cannot be transferred to other backgrounds.In response to the above problems,this topic proposes a new camouflage design method based on Generative Adversarial Networks,which generates camouflage image for a specified detection network to achieve targeted deception of the detection network.This topic takes the characteristics of remote sensing images and detection networks as the research background,and conducts research around the design method of camouflage image of the targets in remote sensing images.We focus on the detection network,generative adversarial network,and style transfer network technology in deep learning.Based on the existing research,we make some improvements and innovations,and design a camouflage image generation system for remote sensing imagesThis subject completes the process of generating camouflage image by Generative Adversarial Networks.In recent years,attacking neural networks by generative adversarial methods has achieved good results,and this topic has introduced them.Aiming at the versatility of camouflage image in multiple backgrounds,we adopt a method based on style transfer network to achieve the fusion of multiple backgrounds,which improves the universality of camouflage image.In view of the characteristics of large-scale changes in remote sensing image targets and relatively blurry textures,this subject has designed convolutional layer noise fusion and multi-category heat map guidance to improve the accuracy of the detection network.In extracting feature layers,introduce noise fusion to improve the robustness of the detection network and improve the performance of the network on blurry textured targets.At the same time,the heat map label is used to guide the network directional fitting to improve the fitting speed and effect.In view of the fact that the generation process of the camouflage image is too long and the degree of refinement is high,this topic designs the structure and loss function of the generation network.It adopts the design of the separate position and category loss and the layer jump connection,which reduces the number of iterations and improves accuracy of generation.It avoids the situation that the single training time is too long and the generation effect is not good,so that we can improve the generation speed and effect of the camouflage image.In view of the versatility of camouflage image,the background fusion scheme of style transfer is adopted,and the optimal background fusion strategy is adopted to effectively expand the target background types and reduce the dependence of specific background and target.