SAR Image Recognition Based On Deep Learning And Target Feature Fusion

Artificial intelligence is a hot topic in today's society and in science and technology research,and image recognition is an important field of artificial intelligence.SAR image recognition problem is chosen as the research direction in this paper,and the image recognition method based on deep learning and target feature fusion is discussed.At present,flexibility,intelligence and versatility of intelligent algorithms designed with deep learning are constantly challenging the limits of human cognition,and these characteristics are necessary for the development and progress of SAR target recognition technology at present.Therefore,based on the theory of deep learning,it is of great theoretical significance and practical value to explore the key technologies and problems of SAR image recognition.The main work of this paper is as follows:The principle of Gabor filter,LBP operator and HOG operator is introduced.In view of the performance of these three traditional operators,the method of combining traditional operators with convolutional neural network is discussed.Firstly,the above operators are used to process the image,then the two-dimensional feature matrix is input into the convolutional neural network for training.Finally,the framework of the convolutional neural network adapt to the problems studied in this paper is proposed.The simulation results show that classification performance of three operators combined with convolutional neural network is better than the convolutional neural network,and the advantages of the three operators are taken into account.At the same time,its performance is obviously better than three operators without convolutional neural network,and the advantages of convolutional neural network are taken into account.In this paper,considering the insufficient classification performance of the last layer in convolutional neural network,the output of fully connected layer in convolutional neural network is taken out separately,and the convolutional neural network is regarded as a feature extraction algorithm,which uses the traditional classifier with stronger nonlinear mapping ability to classify.At the same time,considering the advantages of extreme learning machine,it is applied to feature classifier,and the algorithm is improved appropriately in view of the shortcomings of the performance of the ordinary extreme learning machine.The results show that the improved extreme learning machine classifier achieves satisfactory results and outperforms the conventional extreme learning machine.This paper presents two SAR image recognition algorithms for deep learning and target feature fusion.The First is feature level fusion,and it uses autoencoder to reduce and fuse the features extracted by the three operators combined with convolutional neural network;The second is decision level fusion,and it uses the improved extreme learning machine to classify the features extracted by the three operators,then uses DS evidence theory to fuse the results of the classification.Simulation results show that both fusion methods are superior to the traditional single feature method.