SAR Image Target Recognition Based on Non-Local Operation

Synthetic aperture radar (SAR) target recognition is an important part of SAR image interpretation. It has been widely used in the field of national defense and national economy. At present, the feature extraction based on convolution is a local operation in space and time. The convolution kernel extracts the features in the local region with a certain step size. The global information of the picture can only be obtained by increasing the number of convolution layers. However, this method will not only increase the difficulty of model training but also makes the optimization of the network more difficult, and even leads to over fitting. Therefore, this paper proposes a SAR image target recognition algorithm based on GoogLeNet -NB, which combines accurate and effective GoogLeNet framework and non-local block(NB). By adding NB to GoogLeNet framework, we can capture more context information, enhance the correlation between pixels and regions, and improve the representation ability of the network. In order to verify the effectiveness of NB, NB is added in different positions of GoogLeNet framework for experimental comparison. Finally, MSTAR database is used to verify the algorithm. The experimental results show that the recognition effect of the GoogLeNet -NB algorithm model proposed in this paper is better than the traditional Alexnet algorithm and GoogLeNet algorithm. In the adding NB in different positions of GoogLeNet framework, adding NB in the front position of GoogLeNet can reduce the loss of information in the training process and obtain more global information. Therefore, GoogLeNet-preNB algorithm has certain advantages over GoogLeNet-postNB algorithm.