Convolution neural network SAR image target recognition based on transfer learning

Aiming at the problem of slow convergence and over-fitting in convolutional neural networks due to random initialization and excessive parameters of network parameters, a supervised pre-training convolutional neural network was proposed based on the transfer learning. Firstly, the idea of transfer learning was introduced, and a small-scale dataset was used as a training sample of the source domain. The source task was supervised and trained for the source task. Secondly, a multi-layer convolutional neural network was constructed as the target domain, then the pre-training model obtained from the source domain was taken as the initial parameter of the network, large-scale dataset was used as a training sample for the target domain to fine-tune the network. Through this feature-based transfer learning, the feature information was transfered from source domain to target domain. Aiming at the over-parameterized problem in full connected layer of the convolutional neural network, the full connected layer was replaced by the convolutional layer. Three types of target data in the MSTAR dataset of the US Defense Advanced Research Projects Agency were used as the source domain samples, ten types of target data were used as the target domain samples. The experimental results show that the precision of the ten types of targets is 99.13%, and it has a faster error convergence rate.