Study On Few-shot ISAR Image Recognition Based On Deep Convolutional Neural Networks

Inverse Synthetic Aperture Radar(ISAR)has the characteristics of all-day,all-weather,and high resolution.ISAR plays a significant role in automatic target recognition(ATR)because it can obtain important information such as the shape and structure of the target by highresolution two-dimensional imaging.Generally,traditional ISAR target recognition algorithms have complicated steps,and require manual feature extraction and prior knowledge,thus have low degree of intelligence.In recent years,Deep Convolutional Neural Network(DCNN)has been widely utilized in Synthetic Aperture Radar(SAR)image recognition.DCNN can automatically extract valid features from data without much requirement on manual feature extractor design or much professional knowledge,thus has achieved excellent performance.However,ATR methods based on deep learning require a large number of samples.Usually,ISAR targets are non-cooperative and the observation conditions are restricted,thus one can only obtain limited number of images.In this scenario,DCNN will encounter difficulties such as insufficient model training and fail to obtain good recognition performance.This thesis focuses on few-shot recognition of ISAR images based on DCNN.The specific research contents include: few-shot ISAR target recognition based on prototypical network;few-shot ISAR target recognition based on Gaussian prototypical network with improved metric criteria;few-shot ISAR target recognition based on relational network with automated learned metrics.The relevant studies make full use of few-shot learning and DCNN theory to effectively improve the recognition accuracy in the case of insufficient ISAR images,which provide theoretical and technical support for few-shot ISAR target recognition.The main content of this thesis mainly includes three parts:Aiming at the problem that the current ISAR ATR methods based on DCNN have low recognition accuracy in the case of insufficient samples,the first part discusses the prototypical network for few-shot ISAR target recognition.The method calculates the prototypes for each type of ISAR images by comparing the similarity of their features,and then performs recognition according to the Euclidean distance from the target image to the prototype.This method can achieve high recognition accuracy when the number of ISAR image samples is small.Aiming at the problem that the metric space of the prototypical network is too simple and the extraction of image feature is not sufficient,the second part researches the Gaussian prototypical network for few-shot ISAR target recognition,which increases the trainable parameters of the network and enriches the metric space.Specifically,the Gaussian prototypical network maps the image to the measure vector,predicts its confidence region around,and outputs the embedded vector and matrix atom through the encoder for Gaussian covariance matrix calculation.In this way,the complexity of metric space is increased and higher-level features of the unknown ISAR image are extracted,which will boost the ISAR target recognition performance.Because it is difficult for the manually designed measure to accurately capture the unknown complex image characteristics,there may exist an information loss during feature extraction.To improve the performance of the Gaussian prototypical network,the third part discusses the relation network of few-shot ISAR target recognition,which changes the artificially designed measure to automatically learned measure through network training.In addition,end-to-end learning is realized and more efficient recognition is obtained.