Research On SAR Ship Image Intelligent Recognition Technolog

The ship target recognition in synthetic aperture radar(SAR)image is an important part of SAR image maritime automatic target recognition system.Spaceborne SAR can perform high-resolution imaging all day and all weather,and plays an irreplaceable role in strategic intelligence support and national economic construction.Therefore,it is of great significance to study the method of ship target recognition in SAR image.This paper focuses on the research on the intelligent recognition technology of SAR ship images based on convolutional neural network.According to the existing SAR civil ship data set,the SAR images of bulk cargo ships,container ships and oil tankers are recognized.The main work of this paper is as follows:(1)In view of the SAR imaging mechanism and the characteristics of SAR ship images,the optical image lightweight recognition model MobileNet-V3 is improved,focusing on the use of multi-scale and large-size convolution kernel to extract more feature information of ship target while reducing the amount of network parameters.The experimental results show that compared with the original network,the recognition accuracy on opensarship dataset is increased by 2%,while the scale of network parameters is reduced by 42%;(2)Aiming at the shortage of SAR ship image data sets and the imbalance of categories,the database augmentation method based on Generative Adversarial Network(GAN)is studied.The differentiable data augmentation method is introduced into the GAN network,and the same differentiable augmentation operation is applied to the generated data and real data during training.Experimental results show that this method can generate high-quality realistic images under the condition of few-shot,effectively expand the SAR image data set,alleviate the over fitting phenomenon in few-shot learning,and improve the classification accuracy of SAR images under few-shot condition;(3)On the basis of data augmentation,an overall scheme for fine-grained classification of SAR ship targets based on transfer learning is proposed,and three fine-tuning training strategies are designed to find the most suitable training scheme for SAR ship target recognition.The experimental results show that the bottom fine-tuning training strategy is the best.Using the OpenSarShip data set,the recognition accuracy of resnet 18 and mobilenetv3-small classification models reached 77.54% and 73.64% respectively,which is about 3% higher than that of direct training;(4)The basic idea of meta-learning and the meta-learning domain generalization(MLDG)algorithm are studied,and the MLDG algorithm is used to solve the generalization problem between different SAR ship image datasets.The comparative experimental results show that the method can significantly improve the generalization ability of the SAR ship image recognition network under the condition of few-shot.