Research And Application Of Remote Sensing Ship Data Augmentation Method Based On Deep Learning And Simulation Images

In recent years,remote sensing imaging technology and deep learning methods have developed rapidly.In the field of intelligent processing and interpretation of remote sensing images,deep learning has shown great potential.Sufficient and welllabeled training data is one of the main drivers of deep learning algorithm performance.However,in some non-cooperative scenarios,the scarcity of remote sensing images often leads to low performance of deep learning models.The simulation framework based on imaging simulation technology can generate a large amount of image data flexibly and conveniently.Therefore,in this paper,aiming at the scarcity of real images in the task of remote sensing ship image classification,the research and application of remote sensing ship data augmentation methods based on simulated images are carried out.Since simulated ship images cannot simulate all the detailed features of real ship images,it is necessary to analyze the feasibility of applying simulated images to deep learning.This paper proposes a data augmentation method using simulated images as training samples.The method constructs a simulated ship image dataset by building a simulation framework of typical ship targets in the sea scene,and forms a mixed dataset with a small number of real ship images,which is used as the training data of the neural network to carry out iterative training of the classification model.The experimental results show that compared with the mainstream data augmentation methods,this method can further improve the accuracy of the remote sensing ship image classification model by 1%～2%,which fully verifies the effectiveness of the method and also illustrates the application of simulated images.Feasibility of deep learning.Aiming at the problem of domain differences between simulated ship images and real ship images in terms of style features,this paper proposes a progressive data augmentation method based on simulated images and style transfer networks.The method uses the improved style transfer network to achieve the style feature alignment from the simulated ship image to the real ship image,and then the transferred ship image and a small number of real ship images form a mixed dataset,which is used as a neural network classification model.training data for iterative training.The experimental results show that compared with the mainstream data augmentation methods and the data augmentation method based on simulated images proposed in this paper,this method can further improve the accuracy of the remote sensing ship image classification model by 2%～3%,which fully verifies the method.At the same time,it pointed out the research direction of data augmentation of deep learning models based on simulated images.Finally,combined with the actual needs of data security and technical autonomy,the remote sensing ship image classification algorithm and the progressive data augmentation method proposed in this paper are transplanted and deployed in the domestic AI chip Cambrian MLU220,and a Cambrian based AI chip is built.The remote sensing ship data amplification and identification system of Ji MLU220 has tested and analyzed the overall performance indicators of the system.The results show that the entire system is running well,and the accuracy of remote sensing ship image classification is 86.58%,the frame rate can reach 32 frames to meet the real-time operation requirements.