Research On Target Recognition Of SAR Image Based On Deep Learning

Synthetic Aperture Radar(SAR)is a kind of active ground observation system based on aircraft,satellite,spacecraft and other platforms,and it can observe the ground all day long.With the continuous progress and improvement of SAR system and imaging technology,a large number of SAR images need to be processed,it has become a research trend in the field of SAR image interpretation to obtain feature information from many images based on complex scenes and apply it to target classification and recognition.SAR image target recognition research is mainly based on feature extraction.The features extracted by traditional object recognition methods are mostly based on artificial design,which requires a lot of experimental basis and professional knowledge,and these characteristics are based on the low-level visual features of the target,it is unable to fully represent the essence of the target attribute,cannot obtain useful feature in many cases,lead to poor performance of target recognition.Deep learning as a machine learning method based on the data were characterized learning,can effectively overcome the limitations of traditional image target recognition method,has become a research hotspot in recent years.In this paper,SAR image target recognition is studied based on the deep learning method.The main work is as follows:Aiming at the problem of target recognition of SAR images,an improved image classification algorithm based on convolution neural network is proposed.In order to overcome the problem of over-fitting due to insufficient data in the training process of convolutional neural network,data enhancement is adopted to increase the size of training samples.A multi-scale convolution module is used to replace the high-level convolutional layer.In the output layer,the combination of convolution and global mean pooling is used to replace the traditional full connection layer,which greatly reduces network parameters.During the network training phase,the network parameters are updated by the error back propagation.According to the target and scene classification of the MSTAR dataset and the high resolution airborne SAR images,the experimental results show that the algorithm achieves better classification performance.To solve the problem of slow convergence and over-fitting in convolutional neural networks due to random initialization and excessive parameters of network parameters,a convolution neural network based on transfer learning and supervised pre-training is proposed.First,the idea of transfer learning is introduced,and a small data set is used as the training sample of the source domain.The pre-training model is obtained by conducting supervised training for the source task in the source domain.Then,to build a multi-level convolution neural network as the target domain goal task for training the network,the source domain to obtain the training model as the initial parameters of the network,large-scale data as training samples of target domain network of fine-tuning.Through this kind of transfer learning based on feature selection,the transfer of feature knowledge from source domain to target domain is realized.In order to solve the problem of too many parameters of total connection layer in convolution neural network,the convolution layer is used instead of total connection layer.The experimental results show that the algorithm effectively improves the convergence speed and recognition accuracy.